1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
13 #include "mds_client.h"
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/messenger.h>
19 * Capability management
21 * The Ceph metadata servers control client access to inode metadata
22 * and file data by issuing capabilities, granting clients permission
23 * to read and/or write both inode field and file data to OSDs
24 * (storage nodes). Each capability consists of a set of bits
25 * indicating which operations are allowed.
27 * If the client holds a *_SHARED cap, the client has a coherent value
28 * that can be safely read from the cached inode.
30 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
31 * client is allowed to change inode attributes (e.g., file size,
32 * mtime), note its dirty state in the ceph_cap, and asynchronously
33 * flush that metadata change to the MDS.
35 * In the event of a conflicting operation (perhaps by another
36 * client), the MDS will revoke the conflicting client capabilities.
38 * In order for a client to cache an inode, it must hold a capability
39 * with at least one MDS server. When inodes are released, release
40 * notifications are batched and periodically sent en masse to the MDS
41 * cluster to release server state.
44 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
45 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
46 struct ceph_mds_session *session,
47 struct ceph_inode_info *ci,
48 u64 oldest_flush_tid);
51 * Generate readable cap strings for debugging output.
53 #define MAX_CAP_STR 20
54 static char cap_str[MAX_CAP_STR][40];
55 static DEFINE_SPINLOCK(cap_str_lock);
56 static int last_cap_str;
58 static char *gcap_string(char *s, int c)
60 if (c & CEPH_CAP_GSHARED)
62 if (c & CEPH_CAP_GEXCL)
64 if (c & CEPH_CAP_GCACHE)
70 if (c & CEPH_CAP_GBUFFER)
72 if (c & CEPH_CAP_GLAZYIO)
77 const char *ceph_cap_string(int caps)
83 spin_lock(&cap_str_lock);
85 if (last_cap_str == MAX_CAP_STR)
87 spin_unlock(&cap_str_lock);
91 if (caps & CEPH_CAP_PIN)
94 c = (caps >> CEPH_CAP_SAUTH) & 3;
97 s = gcap_string(s, c);
100 c = (caps >> CEPH_CAP_SLINK) & 3;
103 s = gcap_string(s, c);
106 c = (caps >> CEPH_CAP_SXATTR) & 3;
109 s = gcap_string(s, c);
112 c = caps >> CEPH_CAP_SFILE;
115 s = gcap_string(s, c);
124 void ceph_caps_init(struct ceph_mds_client *mdsc)
126 INIT_LIST_HEAD(&mdsc->caps_list);
127 spin_lock_init(&mdsc->caps_list_lock);
130 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
132 struct ceph_cap *cap;
134 spin_lock(&mdsc->caps_list_lock);
135 while (!list_empty(&mdsc->caps_list)) {
136 cap = list_first_entry(&mdsc->caps_list,
137 struct ceph_cap, caps_item);
138 list_del(&cap->caps_item);
139 kmem_cache_free(ceph_cap_cachep, cap);
141 mdsc->caps_total_count = 0;
142 mdsc->caps_avail_count = 0;
143 mdsc->caps_use_count = 0;
144 mdsc->caps_reserve_count = 0;
145 mdsc->caps_min_count = 0;
146 spin_unlock(&mdsc->caps_list_lock);
149 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
151 spin_lock(&mdsc->caps_list_lock);
152 mdsc->caps_min_count += delta;
153 BUG_ON(mdsc->caps_min_count < 0);
154 spin_unlock(&mdsc->caps_list_lock);
157 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
158 struct ceph_cap_reservation *ctx, int need)
161 struct ceph_cap *cap;
166 dout("reserve caps ctx=%p need=%d\n", ctx, need);
168 /* first reserve any caps that are already allocated */
169 spin_lock(&mdsc->caps_list_lock);
170 if (mdsc->caps_avail_count >= need)
173 have = mdsc->caps_avail_count;
174 mdsc->caps_avail_count -= have;
175 mdsc->caps_reserve_count += have;
176 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
177 mdsc->caps_reserve_count +
178 mdsc->caps_avail_count);
179 spin_unlock(&mdsc->caps_list_lock);
181 for (i = have; i < need; i++) {
182 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
185 list_add(&cap->caps_item, &newcaps);
188 /* we didn't manage to reserve as much as we needed */
189 if (have + alloc != need)
190 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
191 ctx, need, have + alloc);
193 spin_lock(&mdsc->caps_list_lock);
194 mdsc->caps_total_count += alloc;
195 mdsc->caps_reserve_count += alloc;
196 list_splice(&newcaps, &mdsc->caps_list);
198 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
199 mdsc->caps_reserve_count +
200 mdsc->caps_avail_count);
201 spin_unlock(&mdsc->caps_list_lock);
204 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
205 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
206 mdsc->caps_reserve_count, mdsc->caps_avail_count);
209 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
210 struct ceph_cap_reservation *ctx)
212 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
214 spin_lock(&mdsc->caps_list_lock);
215 BUG_ON(mdsc->caps_reserve_count < ctx->count);
216 mdsc->caps_reserve_count -= ctx->count;
217 mdsc->caps_avail_count += ctx->count;
219 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
220 mdsc->caps_total_count, mdsc->caps_use_count,
221 mdsc->caps_reserve_count, mdsc->caps_avail_count);
222 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
223 mdsc->caps_reserve_count +
224 mdsc->caps_avail_count);
225 spin_unlock(&mdsc->caps_list_lock);
230 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
231 struct ceph_cap_reservation *ctx)
233 struct ceph_cap *cap = NULL;
235 /* temporary, until we do something about cap import/export */
237 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
239 spin_lock(&mdsc->caps_list_lock);
240 mdsc->caps_use_count++;
241 mdsc->caps_total_count++;
242 spin_unlock(&mdsc->caps_list_lock);
247 spin_lock(&mdsc->caps_list_lock);
248 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
249 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
250 mdsc->caps_reserve_count, mdsc->caps_avail_count);
252 BUG_ON(ctx->count > mdsc->caps_reserve_count);
253 BUG_ON(list_empty(&mdsc->caps_list));
256 mdsc->caps_reserve_count--;
257 mdsc->caps_use_count++;
259 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
260 list_del(&cap->caps_item);
262 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
263 mdsc->caps_reserve_count + mdsc->caps_avail_count);
264 spin_unlock(&mdsc->caps_list_lock);
268 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
270 spin_lock(&mdsc->caps_list_lock);
271 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
272 cap, mdsc->caps_total_count, mdsc->caps_use_count,
273 mdsc->caps_reserve_count, mdsc->caps_avail_count);
274 mdsc->caps_use_count--;
276 * Keep some preallocated caps around (ceph_min_count), to
277 * avoid lots of free/alloc churn.
279 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
280 mdsc->caps_min_count) {
281 mdsc->caps_total_count--;
282 kmem_cache_free(ceph_cap_cachep, cap);
284 mdsc->caps_avail_count++;
285 list_add(&cap->caps_item, &mdsc->caps_list);
288 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
289 mdsc->caps_reserve_count + mdsc->caps_avail_count);
290 spin_unlock(&mdsc->caps_list_lock);
293 void ceph_reservation_status(struct ceph_fs_client *fsc,
294 int *total, int *avail, int *used, int *reserved,
297 struct ceph_mds_client *mdsc = fsc->mdsc;
300 *total = mdsc->caps_total_count;
302 *avail = mdsc->caps_avail_count;
304 *used = mdsc->caps_use_count;
306 *reserved = mdsc->caps_reserve_count;
308 *min = mdsc->caps_min_count;
312 * Find ceph_cap for given mds, if any.
314 * Called with i_ceph_lock held.
316 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
318 struct ceph_cap *cap;
319 struct rb_node *n = ci->i_caps.rb_node;
322 cap = rb_entry(n, struct ceph_cap, ci_node);
325 else if (mds > cap->mds)
333 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
335 struct ceph_cap *cap;
337 spin_lock(&ci->i_ceph_lock);
338 cap = __get_cap_for_mds(ci, mds);
339 spin_unlock(&ci->i_ceph_lock);
344 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
346 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
348 struct ceph_cap *cap;
352 /* prefer mds with WR|BUFFER|EXCL caps */
353 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
354 cap = rb_entry(p, struct ceph_cap, ci_node);
356 if (cap->issued & (CEPH_CAP_FILE_WR |
357 CEPH_CAP_FILE_BUFFER |
364 int ceph_get_cap_mds(struct inode *inode)
366 struct ceph_inode_info *ci = ceph_inode(inode);
368 spin_lock(&ci->i_ceph_lock);
369 mds = __ceph_get_cap_mds(ceph_inode(inode));
370 spin_unlock(&ci->i_ceph_lock);
375 * Called under i_ceph_lock.
377 static void __insert_cap_node(struct ceph_inode_info *ci,
378 struct ceph_cap *new)
380 struct rb_node **p = &ci->i_caps.rb_node;
381 struct rb_node *parent = NULL;
382 struct ceph_cap *cap = NULL;
386 cap = rb_entry(parent, struct ceph_cap, ci_node);
387 if (new->mds < cap->mds)
389 else if (new->mds > cap->mds)
395 rb_link_node(&new->ci_node, parent, p);
396 rb_insert_color(&new->ci_node, &ci->i_caps);
400 * (re)set cap hold timeouts, which control the delayed release
401 * of unused caps back to the MDS. Should be called on cap use.
403 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
404 struct ceph_inode_info *ci)
406 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
408 ci->i_hold_caps_min = round_jiffies(jiffies +
409 ma->caps_wanted_delay_min * HZ);
410 ci->i_hold_caps_max = round_jiffies(jiffies +
411 ma->caps_wanted_delay_max * HZ);
412 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
413 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
417 * (Re)queue cap at the end of the delayed cap release list.
419 * If I_FLUSH is set, leave the inode at the front of the list.
421 * Caller holds i_ceph_lock
422 * -> we take mdsc->cap_delay_lock
424 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
425 struct ceph_inode_info *ci)
427 __cap_set_timeouts(mdsc, ci);
428 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
429 ci->i_ceph_flags, ci->i_hold_caps_max);
430 if (!mdsc->stopping) {
431 spin_lock(&mdsc->cap_delay_lock);
432 if (!list_empty(&ci->i_cap_delay_list)) {
433 if (ci->i_ceph_flags & CEPH_I_FLUSH)
435 list_del_init(&ci->i_cap_delay_list);
437 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
439 spin_unlock(&mdsc->cap_delay_lock);
444 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
445 * indicating we should send a cap message to flush dirty metadata
446 * asap, and move to the front of the delayed cap list.
448 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
449 struct ceph_inode_info *ci)
451 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
452 spin_lock(&mdsc->cap_delay_lock);
453 ci->i_ceph_flags |= CEPH_I_FLUSH;
454 if (!list_empty(&ci->i_cap_delay_list))
455 list_del_init(&ci->i_cap_delay_list);
456 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
457 spin_unlock(&mdsc->cap_delay_lock);
461 * Cancel delayed work on cap.
463 * Caller must hold i_ceph_lock.
465 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
466 struct ceph_inode_info *ci)
468 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
469 if (list_empty(&ci->i_cap_delay_list))
471 spin_lock(&mdsc->cap_delay_lock);
472 list_del_init(&ci->i_cap_delay_list);
473 spin_unlock(&mdsc->cap_delay_lock);
477 * Common issue checks for add_cap, handle_cap_grant.
479 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
482 unsigned had = __ceph_caps_issued(ci, NULL);
485 * Each time we receive FILE_CACHE anew, we increment
488 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
489 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
494 * If FILE_SHARED is newly issued, mark dir not complete. We don't
495 * know what happened to this directory while we didn't have the cap.
496 * If FILE_SHARED is being revoked, also mark dir not complete. It
497 * stops on-going cached readdir.
499 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
500 if (issued & CEPH_CAP_FILE_SHARED)
502 if (S_ISDIR(ci->vfs_inode.i_mode)) {
503 dout(" marking %p NOT complete\n", &ci->vfs_inode);
504 __ceph_dir_clear_complete(ci);
510 * Add a capability under the given MDS session.
512 * Caller should hold session snap_rwsem (read) and s_mutex.
514 * @fmode is the open file mode, if we are opening a file, otherwise
515 * it is < 0. (This is so we can atomically add the cap and add an
516 * open file reference to it.)
518 void ceph_add_cap(struct inode *inode,
519 struct ceph_mds_session *session, u64 cap_id,
520 int fmode, unsigned issued, unsigned wanted,
521 unsigned seq, unsigned mseq, u64 realmino, int flags,
522 struct ceph_cap **new_cap)
524 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
525 struct ceph_inode_info *ci = ceph_inode(inode);
526 struct ceph_cap *cap;
527 int mds = session->s_mds;
530 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
531 session->s_mds, cap_id, ceph_cap_string(issued), seq);
534 * If we are opening the file, include file mode wanted bits
538 wanted |= ceph_caps_for_mode(fmode);
540 cap = __get_cap_for_mds(ci, mds);
546 cap->implemented = 0;
552 __insert_cap_node(ci, cap);
554 /* add to session cap list */
555 cap->session = session;
556 spin_lock(&session->s_cap_lock);
557 list_add_tail(&cap->session_caps, &session->s_caps);
558 session->s_nr_caps++;
559 spin_unlock(&session->s_cap_lock);
562 * auth mds of the inode changed. we received the cap export
563 * message, but still haven't received the cap import message.
564 * handle_cap_export() updated the new auth MDS' cap.
566 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
567 * a message that was send before the cap import message. So
570 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
571 WARN_ON(cap != ci->i_auth_cap);
572 WARN_ON(cap->cap_id != cap_id);
575 issued |= cap->issued;
576 flags |= CEPH_CAP_FLAG_AUTH;
580 if (!ci->i_snap_realm) {
582 * add this inode to the appropriate snap realm
584 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
587 spin_lock(&realm->inodes_with_caps_lock);
588 ci->i_snap_realm = realm;
589 list_add(&ci->i_snap_realm_item,
590 &realm->inodes_with_caps);
591 spin_unlock(&realm->inodes_with_caps_lock);
593 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
599 __check_cap_issue(ci, cap, issued);
602 * If we are issued caps we don't want, or the mds' wanted
603 * value appears to be off, queue a check so we'll release
604 * later and/or update the mds wanted value.
606 actual_wanted = __ceph_caps_wanted(ci);
607 if ((wanted & ~actual_wanted) ||
608 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
609 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
610 ceph_cap_string(issued), ceph_cap_string(wanted),
611 ceph_cap_string(actual_wanted));
612 __cap_delay_requeue(mdsc, ci);
615 if (flags & CEPH_CAP_FLAG_AUTH) {
616 if (!ci->i_auth_cap ||
617 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
618 ci->i_auth_cap = cap;
619 cap->mds_wanted = wanted;
622 WARN_ON(ci->i_auth_cap == cap);
625 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
626 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
627 ceph_cap_string(issued|cap->issued), seq, mds);
628 cap->cap_id = cap_id;
629 cap->issued = issued;
630 cap->implemented |= issued;
631 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
632 cap->mds_wanted = wanted;
634 cap->mds_wanted |= wanted;
636 cap->issue_seq = seq;
638 cap->cap_gen = session->s_cap_gen;
641 __ceph_get_fmode(ci, fmode);
645 * Return true if cap has not timed out and belongs to the current
646 * generation of the MDS session (i.e. has not gone 'stale' due to
647 * us losing touch with the mds).
649 static int __cap_is_valid(struct ceph_cap *cap)
654 spin_lock(&cap->session->s_gen_ttl_lock);
655 gen = cap->session->s_cap_gen;
656 ttl = cap->session->s_cap_ttl;
657 spin_unlock(&cap->session->s_gen_ttl_lock);
659 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
660 dout("__cap_is_valid %p cap %p issued %s "
661 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
662 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
670 * Return set of valid cap bits issued to us. Note that caps time
671 * out, and may be invalidated in bulk if the client session times out
672 * and session->s_cap_gen is bumped.
674 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
676 int have = ci->i_snap_caps;
677 struct ceph_cap *cap;
682 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
683 cap = rb_entry(p, struct ceph_cap, ci_node);
684 if (!__cap_is_valid(cap))
686 dout("__ceph_caps_issued %p cap %p issued %s\n",
687 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
690 *implemented |= cap->implemented;
693 * exclude caps issued by non-auth MDS, but are been revoking
694 * by the auth MDS. The non-auth MDS should be revoking/exporting
695 * these caps, but the message is delayed.
697 if (ci->i_auth_cap) {
698 cap = ci->i_auth_cap;
699 have &= ~cap->implemented | cap->issued;
705 * Get cap bits issued by caps other than @ocap
707 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
709 int have = ci->i_snap_caps;
710 struct ceph_cap *cap;
713 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
714 cap = rb_entry(p, struct ceph_cap, ci_node);
717 if (!__cap_is_valid(cap))
725 * Move a cap to the end of the LRU (oldest caps at list head, newest
728 static void __touch_cap(struct ceph_cap *cap)
730 struct ceph_mds_session *s = cap->session;
732 spin_lock(&s->s_cap_lock);
733 if (!s->s_cap_iterator) {
734 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
736 list_move_tail(&cap->session_caps, &s->s_caps);
738 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
739 &cap->ci->vfs_inode, cap, s->s_mds);
741 spin_unlock(&s->s_cap_lock);
745 * Check if we hold the given mask. If so, move the cap(s) to the
746 * front of their respective LRUs. (This is the preferred way for
747 * callers to check for caps they want.)
749 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
751 struct ceph_cap *cap;
753 int have = ci->i_snap_caps;
755 if ((have & mask) == mask) {
756 dout("__ceph_caps_issued_mask %p snap issued %s"
757 " (mask %s)\n", &ci->vfs_inode,
758 ceph_cap_string(have),
759 ceph_cap_string(mask));
763 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
764 cap = rb_entry(p, struct ceph_cap, ci_node);
765 if (!__cap_is_valid(cap))
767 if ((cap->issued & mask) == mask) {
768 dout("__ceph_caps_issued_mask %p cap %p issued %s"
769 " (mask %s)\n", &ci->vfs_inode, cap,
770 ceph_cap_string(cap->issued),
771 ceph_cap_string(mask));
777 /* does a combination of caps satisfy mask? */
779 if ((have & mask) == mask) {
780 dout("__ceph_caps_issued_mask %p combo issued %s"
781 " (mask %s)\n", &ci->vfs_inode,
782 ceph_cap_string(cap->issued),
783 ceph_cap_string(mask));
787 /* touch this + preceding caps */
789 for (q = rb_first(&ci->i_caps); q != p;
791 cap = rb_entry(q, struct ceph_cap,
793 if (!__cap_is_valid(cap))
806 * Return true if mask caps are currently being revoked by an MDS.
808 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
809 struct ceph_cap *ocap, int mask)
811 struct ceph_cap *cap;
814 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
815 cap = rb_entry(p, struct ceph_cap, ci_node);
817 (cap->implemented & ~cap->issued & mask))
823 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
825 struct inode *inode = &ci->vfs_inode;
828 spin_lock(&ci->i_ceph_lock);
829 ret = __ceph_caps_revoking_other(ci, NULL, mask);
830 spin_unlock(&ci->i_ceph_lock);
831 dout("ceph_caps_revoking %p %s = %d\n", inode,
832 ceph_cap_string(mask), ret);
836 int __ceph_caps_used(struct ceph_inode_info *ci)
840 used |= CEPH_CAP_PIN;
842 used |= CEPH_CAP_FILE_RD;
843 if (ci->i_rdcache_ref ||
844 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
845 ci->vfs_inode.i_data.nrpages))
846 used |= CEPH_CAP_FILE_CACHE;
848 used |= CEPH_CAP_FILE_WR;
849 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
850 used |= CEPH_CAP_FILE_BUFFER;
855 * wanted, by virtue of open file modes
857 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
860 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
861 if (ci->i_nr_by_mode[i])
866 return ceph_caps_for_mode(bits >> 1);
870 * Return caps we have registered with the MDS(s) as 'wanted'.
872 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
874 struct ceph_cap *cap;
878 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
879 cap = rb_entry(p, struct ceph_cap, ci_node);
880 if (check && !__cap_is_valid(cap))
882 if (cap == ci->i_auth_cap)
883 mds_wanted |= cap->mds_wanted;
885 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
891 * called under i_ceph_lock
893 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
895 return !RB_EMPTY_ROOT(&ci->i_caps);
898 int ceph_is_any_caps(struct inode *inode)
900 struct ceph_inode_info *ci = ceph_inode(inode);
903 spin_lock(&ci->i_ceph_lock);
904 ret = __ceph_is_any_caps(ci);
905 spin_unlock(&ci->i_ceph_lock);
910 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
912 struct ceph_snap_realm *realm = ci->i_snap_realm;
913 spin_lock(&realm->inodes_with_caps_lock);
914 list_del_init(&ci->i_snap_realm_item);
915 ci->i_snap_realm_counter++;
916 ci->i_snap_realm = NULL;
917 spin_unlock(&realm->inodes_with_caps_lock);
918 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
923 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
925 * caller should hold i_ceph_lock.
926 * caller will not hold session s_mutex if called from destroy_inode.
928 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
930 struct ceph_mds_session *session = cap->session;
931 struct ceph_inode_info *ci = cap->ci;
932 struct ceph_mds_client *mdsc =
933 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
936 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
938 /* remove from session list */
939 spin_lock(&session->s_cap_lock);
940 if (session->s_cap_iterator == cap) {
941 /* not yet, we are iterating over this very cap */
942 dout("__ceph_remove_cap delaying %p removal from session %p\n",
945 list_del_init(&cap->session_caps);
946 session->s_nr_caps--;
950 /* protect backpointer with s_cap_lock: see iterate_session_caps */
954 * s_cap_reconnect is protected by s_cap_lock. no one changes
955 * s_cap_gen while session is in the reconnect state.
958 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
959 cap->queue_release = 1;
961 list_add_tail(&cap->session_caps,
962 &session->s_cap_releases);
963 session->s_num_cap_releases++;
967 cap->queue_release = 0;
969 cap->cap_ino = ci->i_vino.ino;
971 spin_unlock(&session->s_cap_lock);
973 /* remove from inode list */
974 rb_erase(&cap->ci_node, &ci->i_caps);
975 if (ci->i_auth_cap == cap)
976 ci->i_auth_cap = NULL;
979 ceph_put_cap(mdsc, cap);
981 /* when reconnect denied, we remove session caps forcibly,
982 * i_wr_ref can be non-zero. If there are ongoing write,
985 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
986 drop_inode_snap_realm(ci);
988 if (!__ceph_is_any_real_caps(ci))
989 __cap_delay_cancel(mdsc, ci);
992 struct cap_msg_args {
993 struct ceph_mds_session *session;
994 u64 ino, cid, follows;
995 u64 flush_tid, oldest_flush_tid, size, max_size;
997 struct ceph_buffer *xattr_buf;
998 struct timespec atime, mtime, ctime;
999 int op, caps, wanted, dirty;
1000 u32 seq, issue_seq, mseq, time_warp_seq;
1009 * Build and send a cap message to the given MDS.
1011 * Caller should be holding s_mutex.
1013 static int send_cap_msg(struct cap_msg_args *arg)
1015 struct ceph_mds_caps *fc;
1016 struct ceph_msg *msg;
1019 struct timespec zerotime = {0};
1020 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1022 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1023 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1024 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1025 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1026 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1027 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1028 arg->mseq, arg->follows, arg->size, arg->max_size,
1030 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1032 /* flock buffer size + inline version + inline data size +
1033 * osd_epoch_barrier + oldest_flush_tid */
1034 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1035 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1040 msg->hdr.version = cpu_to_le16(10);
1041 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1043 fc = msg->front.iov_base;
1044 memset(fc, 0, sizeof(*fc));
1046 fc->cap_id = cpu_to_le64(arg->cid);
1047 fc->op = cpu_to_le32(arg->op);
1048 fc->seq = cpu_to_le32(arg->seq);
1049 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1050 fc->migrate_seq = cpu_to_le32(arg->mseq);
1051 fc->caps = cpu_to_le32(arg->caps);
1052 fc->wanted = cpu_to_le32(arg->wanted);
1053 fc->dirty = cpu_to_le32(arg->dirty);
1054 fc->ino = cpu_to_le64(arg->ino);
1055 fc->snap_follows = cpu_to_le64(arg->follows);
1057 fc->size = cpu_to_le64(arg->size);
1058 fc->max_size = cpu_to_le64(arg->max_size);
1059 ceph_encode_timespec(&fc->mtime, &arg->mtime);
1060 ceph_encode_timespec(&fc->atime, &arg->atime);
1061 ceph_encode_timespec(&fc->ctime, &arg->ctime);
1062 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1064 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1065 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1066 fc->mode = cpu_to_le32(arg->mode);
1068 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1069 if (arg->xattr_buf) {
1070 msg->middle = ceph_buffer_get(arg->xattr_buf);
1071 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1072 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1076 /* flock buffer size (version 2) */
1077 ceph_encode_32(&p, 0);
1078 /* inline version (version 4) */
1079 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1080 /* inline data size */
1081 ceph_encode_32(&p, 0);
1083 * osd_epoch_barrier (version 5)
1084 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1085 * case it was recently changed
1087 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1088 /* oldest_flush_tid (version 6) */
1089 ceph_encode_64(&p, arg->oldest_flush_tid);
1092 * caller_uid/caller_gid (version 7)
1094 * Currently, we don't properly track which caller dirtied the caps
1095 * last, and force a flush of them when there is a conflict. For now,
1096 * just set this to 0:0, to emulate how the MDS has worked up to now.
1098 ceph_encode_32(&p, 0);
1099 ceph_encode_32(&p, 0);
1101 /* pool namespace (version 8) (mds always ignores this) */
1102 ceph_encode_32(&p, 0);
1105 * btime and change_attr (version 9)
1107 * We just zero these out for now, as the MDS ignores them unless
1108 * the requisite feature flags are set (which we don't do yet).
1110 ceph_encode_timespec(p, &zerotime);
1111 p += sizeof(struct ceph_timespec);
1112 ceph_encode_64(&p, 0);
1114 /* Advisory flags (version 10) */
1115 ceph_encode_32(&p, arg->flags);
1117 ceph_con_send(&arg->session->s_con, msg);
1122 * Queue cap releases when an inode is dropped from our cache. Since
1123 * inode is about to be destroyed, there is no need for i_ceph_lock.
1125 void ceph_queue_caps_release(struct inode *inode)
1127 struct ceph_inode_info *ci = ceph_inode(inode);
1130 p = rb_first(&ci->i_caps);
1132 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1134 __ceph_remove_cap(cap, true);
1139 * Send a cap msg on the given inode. Update our caps state, then
1140 * drop i_ceph_lock and send the message.
1142 * Make note of max_size reported/requested from mds, revoked caps
1143 * that have now been implemented.
1145 * Make half-hearted attempt ot to invalidate page cache if we are
1146 * dropping RDCACHE. Note that this will leave behind locked pages
1147 * that we'll then need to deal with elsewhere.
1149 * Return non-zero if delayed release, or we experienced an error
1150 * such that the caller should requeue + retry later.
1152 * called with i_ceph_lock, then drops it.
1153 * caller should hold snap_rwsem (read), s_mutex.
1155 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1156 int op, bool sync, int used, int want, int retain,
1157 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1158 __releases(cap->ci->i_ceph_lock)
1160 struct ceph_inode_info *ci = cap->ci;
1161 struct inode *inode = &ci->vfs_inode;
1162 struct cap_msg_args arg;
1163 int held, revoking, dropping;
1168 held = cap->issued | cap->implemented;
1169 revoking = cap->implemented & ~cap->issued;
1170 retain &= ~revoking;
1171 dropping = cap->issued & ~retain;
1173 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1174 inode, cap, cap->session,
1175 ceph_cap_string(held), ceph_cap_string(held & retain),
1176 ceph_cap_string(revoking));
1177 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1179 arg.session = cap->session;
1181 /* don't release wanted unless we've waited a bit. */
1182 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1183 time_before(jiffies, ci->i_hold_caps_min)) {
1184 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1185 ceph_cap_string(cap->issued),
1186 ceph_cap_string(cap->issued & retain),
1187 ceph_cap_string(cap->mds_wanted),
1188 ceph_cap_string(want));
1189 want |= cap->mds_wanted;
1190 retain |= cap->issued;
1193 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1194 if (want & ~cap->mds_wanted) {
1195 /* user space may open/close single file frequently.
1196 * This avoids droping mds_wanted immediately after
1197 * requesting new mds_wanted.
1199 __cap_set_timeouts(mdsc, ci);
1202 cap->issued &= retain; /* drop bits we don't want */
1203 if (cap->implemented & ~cap->issued) {
1205 * Wake up any waiters on wanted -> needed transition.
1206 * This is due to the weird transition from buffered
1207 * to sync IO... we need to flush dirty pages _before_
1208 * allowing sync writes to avoid reordering.
1212 cap->implemented &= cap->issued | used;
1213 cap->mds_wanted = want;
1215 arg.ino = ceph_vino(inode).ino;
1216 arg.cid = cap->cap_id;
1217 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1218 arg.flush_tid = flush_tid;
1219 arg.oldest_flush_tid = oldest_flush_tid;
1221 arg.size = inode->i_size;
1222 ci->i_reported_size = arg.size;
1223 arg.max_size = ci->i_wanted_max_size;
1224 ci->i_requested_max_size = arg.max_size;
1226 if (flushing & CEPH_CAP_XATTR_EXCL) {
1227 __ceph_build_xattrs_blob(ci);
1228 arg.xattr_version = ci->i_xattrs.version;
1229 arg.xattr_buf = ci->i_xattrs.blob;
1231 arg.xattr_buf = NULL;
1234 arg.mtime = inode->i_mtime;
1235 arg.atime = inode->i_atime;
1236 arg.ctime = inode->i_ctime;
1239 arg.caps = cap->implemented;
1241 arg.dirty = flushing;
1244 arg.issue_seq = cap->issue_seq;
1245 arg.mseq = cap->mseq;
1246 arg.time_warp_seq = ci->i_time_warp_seq;
1248 arg.uid = inode->i_uid;
1249 arg.gid = inode->i_gid;
1250 arg.mode = inode->i_mode;
1252 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1253 if (list_empty(&ci->i_cap_snaps))
1254 arg.flags = CEPH_CLIENT_CAPS_NO_CAPSNAP;
1256 arg.flags = CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1258 arg.flags |= CEPH_CLIENT_CAPS_SYNC;
1260 spin_unlock(&ci->i_ceph_lock);
1262 ret = send_cap_msg(&arg);
1264 dout("error sending cap msg, must requeue %p\n", inode);
1269 wake_up_all(&ci->i_cap_wq);
1274 static inline int __send_flush_snap(struct inode *inode,
1275 struct ceph_mds_session *session,
1276 struct ceph_cap_snap *capsnap,
1277 u32 mseq, u64 oldest_flush_tid)
1279 struct cap_msg_args arg;
1281 arg.session = session;
1282 arg.ino = ceph_vino(inode).ino;
1284 arg.follows = capsnap->follows;
1285 arg.flush_tid = capsnap->cap_flush.tid;
1286 arg.oldest_flush_tid = oldest_flush_tid;
1288 arg.size = capsnap->size;
1290 arg.xattr_version = capsnap->xattr_version;
1291 arg.xattr_buf = capsnap->xattr_blob;
1293 arg.atime = capsnap->atime;
1294 arg.mtime = capsnap->mtime;
1295 arg.ctime = capsnap->ctime;
1297 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1298 arg.caps = capsnap->issued;
1300 arg.dirty = capsnap->dirty;
1305 arg.time_warp_seq = capsnap->time_warp_seq;
1307 arg.uid = capsnap->uid;
1308 arg.gid = capsnap->gid;
1309 arg.mode = capsnap->mode;
1311 arg.inline_data = capsnap->inline_data;
1314 return send_cap_msg(&arg);
1318 * When a snapshot is taken, clients accumulate dirty metadata on
1319 * inodes with capabilities in ceph_cap_snaps to describe the file
1320 * state at the time the snapshot was taken. This must be flushed
1321 * asynchronously back to the MDS once sync writes complete and dirty
1322 * data is written out.
1324 * Called under i_ceph_lock. Takes s_mutex as needed.
1326 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1327 struct ceph_mds_session *session)
1328 __releases(ci->i_ceph_lock)
1329 __acquires(ci->i_ceph_lock)
1331 struct inode *inode = &ci->vfs_inode;
1332 struct ceph_mds_client *mdsc = session->s_mdsc;
1333 struct ceph_cap_snap *capsnap;
1334 u64 oldest_flush_tid = 0;
1335 u64 first_tid = 1, last_tid = 0;
1337 dout("__flush_snaps %p session %p\n", inode, session);
1339 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1341 * we need to wait for sync writes to complete and for dirty
1342 * pages to be written out.
1344 if (capsnap->dirty_pages || capsnap->writing)
1347 /* should be removed by ceph_try_drop_cap_snap() */
1348 BUG_ON(!capsnap->need_flush);
1350 /* only flush each capsnap once */
1351 if (capsnap->cap_flush.tid > 0) {
1352 dout(" already flushed %p, skipping\n", capsnap);
1356 spin_lock(&mdsc->cap_dirty_lock);
1357 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1358 list_add_tail(&capsnap->cap_flush.g_list,
1359 &mdsc->cap_flush_list);
1360 if (oldest_flush_tid == 0)
1361 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1362 if (list_empty(&ci->i_flushing_item)) {
1363 list_add_tail(&ci->i_flushing_item,
1364 &session->s_cap_flushing);
1366 spin_unlock(&mdsc->cap_dirty_lock);
1368 list_add_tail(&capsnap->cap_flush.i_list,
1369 &ci->i_cap_flush_list);
1372 first_tid = capsnap->cap_flush.tid;
1373 last_tid = capsnap->cap_flush.tid;
1376 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1378 while (first_tid <= last_tid) {
1379 struct ceph_cap *cap = ci->i_auth_cap;
1380 struct ceph_cap_flush *cf;
1383 if (!(cap && cap->session == session)) {
1384 dout("__flush_snaps %p auth cap %p not mds%d, "
1385 "stop\n", inode, cap, session->s_mds);
1390 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1391 if (cf->tid >= first_tid) {
1399 first_tid = cf->tid + 1;
1401 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1402 refcount_inc(&capsnap->nref);
1403 spin_unlock(&ci->i_ceph_lock);
1405 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1406 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1408 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1411 pr_err("__flush_snaps: error sending cap flushsnap, "
1412 "ino (%llx.%llx) tid %llu follows %llu\n",
1413 ceph_vinop(inode), cf->tid, capsnap->follows);
1416 ceph_put_cap_snap(capsnap);
1417 spin_lock(&ci->i_ceph_lock);
1421 void ceph_flush_snaps(struct ceph_inode_info *ci,
1422 struct ceph_mds_session **psession)
1424 struct inode *inode = &ci->vfs_inode;
1425 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1426 struct ceph_mds_session *session = NULL;
1429 dout("ceph_flush_snaps %p\n", inode);
1431 session = *psession;
1433 spin_lock(&ci->i_ceph_lock);
1434 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1435 dout(" no capsnap needs flush, doing nothing\n");
1438 if (!ci->i_auth_cap) {
1439 dout(" no auth cap (migrating?), doing nothing\n");
1443 mds = ci->i_auth_cap->session->s_mds;
1444 if (session && session->s_mds != mds) {
1445 dout(" oops, wrong session %p mutex\n", session);
1446 mutex_unlock(&session->s_mutex);
1447 ceph_put_mds_session(session);
1451 spin_unlock(&ci->i_ceph_lock);
1452 mutex_lock(&mdsc->mutex);
1453 session = __ceph_lookup_mds_session(mdsc, mds);
1454 mutex_unlock(&mdsc->mutex);
1456 dout(" inverting session/ino locks on %p\n", session);
1457 mutex_lock(&session->s_mutex);
1462 // make sure flushsnap messages are sent in proper order.
1463 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1464 __kick_flushing_caps(mdsc, session, ci, 0);
1465 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1468 __ceph_flush_snaps(ci, session);
1470 spin_unlock(&ci->i_ceph_lock);
1473 *psession = session;
1474 } else if (session) {
1475 mutex_unlock(&session->s_mutex);
1476 ceph_put_mds_session(session);
1478 /* we flushed them all; remove this inode from the queue */
1479 spin_lock(&mdsc->snap_flush_lock);
1480 list_del_init(&ci->i_snap_flush_item);
1481 spin_unlock(&mdsc->snap_flush_lock);
1485 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1486 * Caller is then responsible for calling __mark_inode_dirty with the
1487 * returned flags value.
1489 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1490 struct ceph_cap_flush **pcf)
1492 struct ceph_mds_client *mdsc =
1493 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1494 struct inode *inode = &ci->vfs_inode;
1495 int was = ci->i_dirty_caps;
1498 if (!ci->i_auth_cap) {
1499 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1500 "but no auth cap (session was closed?)\n",
1501 inode, ceph_ino(inode), ceph_cap_string(mask));
1505 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1506 ceph_cap_string(mask), ceph_cap_string(was),
1507 ceph_cap_string(was | mask));
1508 ci->i_dirty_caps |= mask;
1510 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1511 swap(ci->i_prealloc_cap_flush, *pcf);
1513 if (!ci->i_head_snapc) {
1514 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1515 ci->i_head_snapc = ceph_get_snap_context(
1516 ci->i_snap_realm->cached_context);
1518 dout(" inode %p now dirty snapc %p auth cap %p\n",
1519 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1520 BUG_ON(!list_empty(&ci->i_dirty_item));
1521 spin_lock(&mdsc->cap_dirty_lock);
1522 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1523 spin_unlock(&mdsc->cap_dirty_lock);
1524 if (ci->i_flushing_caps == 0) {
1526 dirty |= I_DIRTY_SYNC;
1529 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1531 BUG_ON(list_empty(&ci->i_dirty_item));
1532 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1533 (mask & CEPH_CAP_FILE_BUFFER))
1534 dirty |= I_DIRTY_DATASYNC;
1535 __cap_delay_requeue(mdsc, ci);
1539 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1541 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1544 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1547 kmem_cache_free(ceph_cap_flush_cachep, cf);
1550 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1552 if (!list_empty(&mdsc->cap_flush_list)) {
1553 struct ceph_cap_flush *cf =
1554 list_first_entry(&mdsc->cap_flush_list,
1555 struct ceph_cap_flush, g_list);
1562 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1563 * Return true if caller needs to wake up flush waiters.
1565 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1566 struct ceph_inode_info *ci,
1567 struct ceph_cap_flush *cf)
1569 struct ceph_cap_flush *prev;
1570 bool wake = cf->wake;
1572 /* are there older pending cap flushes? */
1573 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1574 prev = list_prev_entry(cf, g_list);
1578 list_del(&cf->g_list);
1580 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1581 prev = list_prev_entry(cf, i_list);
1585 list_del(&cf->i_list);
1593 * Add dirty inode to the flushing list. Assigned a seq number so we
1594 * can wait for caps to flush without starving.
1596 * Called under i_ceph_lock.
1598 static int __mark_caps_flushing(struct inode *inode,
1599 struct ceph_mds_session *session, bool wake,
1600 u64 *flush_tid, u64 *oldest_flush_tid)
1602 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1603 struct ceph_inode_info *ci = ceph_inode(inode);
1604 struct ceph_cap_flush *cf = NULL;
1607 BUG_ON(ci->i_dirty_caps == 0);
1608 BUG_ON(list_empty(&ci->i_dirty_item));
1609 BUG_ON(!ci->i_prealloc_cap_flush);
1611 flushing = ci->i_dirty_caps;
1612 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1613 ceph_cap_string(flushing),
1614 ceph_cap_string(ci->i_flushing_caps),
1615 ceph_cap_string(ci->i_flushing_caps | flushing));
1616 ci->i_flushing_caps |= flushing;
1617 ci->i_dirty_caps = 0;
1618 dout(" inode %p now !dirty\n", inode);
1620 swap(cf, ci->i_prealloc_cap_flush);
1621 cf->caps = flushing;
1624 spin_lock(&mdsc->cap_dirty_lock);
1625 list_del_init(&ci->i_dirty_item);
1627 cf->tid = ++mdsc->last_cap_flush_tid;
1628 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1629 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1631 if (list_empty(&ci->i_flushing_item)) {
1632 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1633 mdsc->num_cap_flushing++;
1635 spin_unlock(&mdsc->cap_dirty_lock);
1637 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1639 *flush_tid = cf->tid;
1644 * try to invalidate mapping pages without blocking.
1646 static int try_nonblocking_invalidate(struct inode *inode)
1648 struct ceph_inode_info *ci = ceph_inode(inode);
1649 u32 invalidating_gen = ci->i_rdcache_gen;
1651 spin_unlock(&ci->i_ceph_lock);
1652 invalidate_mapping_pages(&inode->i_data, 0, -1);
1653 spin_lock(&ci->i_ceph_lock);
1655 if (inode->i_data.nrpages == 0 &&
1656 invalidating_gen == ci->i_rdcache_gen) {
1658 dout("try_nonblocking_invalidate %p success\n", inode);
1659 /* save any racing async invalidate some trouble */
1660 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1663 dout("try_nonblocking_invalidate %p failed\n", inode);
1667 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1669 loff_t size = ci->vfs_inode.i_size;
1670 /* mds will adjust max size according to the reported size */
1671 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1673 if (size >= ci->i_max_size)
1675 /* half of previous max_size increment has been used */
1676 if (ci->i_max_size > ci->i_reported_size &&
1677 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1683 * Swiss army knife function to examine currently used and wanted
1684 * versus held caps. Release, flush, ack revoked caps to mds as
1687 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1688 * cap release further.
1689 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1690 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1693 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1694 struct ceph_mds_session *session)
1696 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1697 struct ceph_mds_client *mdsc = fsc->mdsc;
1698 struct inode *inode = &ci->vfs_inode;
1699 struct ceph_cap *cap;
1700 u64 flush_tid, oldest_flush_tid;
1701 int file_wanted, used, cap_used;
1702 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1703 int issued, implemented, want, retain, revoking, flushing = 0;
1704 int mds = -1; /* keep track of how far we've gone through i_caps list
1705 to avoid an infinite loop on retry */
1707 int delayed = 0, sent = 0, num;
1708 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1709 bool queue_invalidate = false;
1710 bool force_requeue = false;
1711 bool tried_invalidate = false;
1713 /* if we are unmounting, flush any unused caps immediately. */
1717 spin_lock(&ci->i_ceph_lock);
1719 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1720 flags |= CHECK_CAPS_FLUSH;
1724 spin_lock(&ci->i_ceph_lock);
1726 file_wanted = __ceph_caps_file_wanted(ci);
1727 used = __ceph_caps_used(ci);
1728 issued = __ceph_caps_issued(ci, &implemented);
1729 revoking = implemented & ~issued;
1732 retain = file_wanted | used | CEPH_CAP_PIN;
1733 if (!mdsc->stopping && inode->i_nlink > 0) {
1735 retain |= CEPH_CAP_ANY; /* be greedy */
1736 } else if (S_ISDIR(inode->i_mode) &&
1737 (issued & CEPH_CAP_FILE_SHARED) &&
1738 __ceph_dir_is_complete(ci)) {
1740 * If a directory is complete, we want to keep
1741 * the exclusive cap. So that MDS does not end up
1742 * revoking the shared cap on every create/unlink
1745 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1749 retain |= CEPH_CAP_ANY_SHARED;
1751 * keep RD only if we didn't have the file open RW,
1752 * because then the mds would revoke it anyway to
1753 * journal max_size=0.
1755 if (ci->i_max_size == 0)
1756 retain |= CEPH_CAP_ANY_RD;
1760 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1761 " issued %s revoking %s retain %s %s%s%s\n", inode,
1762 ceph_cap_string(file_wanted),
1763 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1764 ceph_cap_string(ci->i_flushing_caps),
1765 ceph_cap_string(issued), ceph_cap_string(revoking),
1766 ceph_cap_string(retain),
1767 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1768 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1769 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1772 * If we no longer need to hold onto old our caps, and we may
1773 * have cached pages, but don't want them, then try to invalidate.
1774 * If we fail, it's because pages are locked.... try again later.
1776 if ((!is_delayed || mdsc->stopping) &&
1777 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1778 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1779 inode->i_data.nrpages && /* have cached pages */
1780 (revoking & (CEPH_CAP_FILE_CACHE|
1781 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1782 !tried_invalidate) {
1783 dout("check_caps trying to invalidate on %p\n", inode);
1784 if (try_nonblocking_invalidate(inode) < 0) {
1785 if (revoking & (CEPH_CAP_FILE_CACHE|
1786 CEPH_CAP_FILE_LAZYIO)) {
1787 dout("check_caps queuing invalidate\n");
1788 queue_invalidate = true;
1789 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1791 dout("check_caps failed to invalidate pages\n");
1792 /* we failed to invalidate pages. check these
1793 caps again later. */
1794 force_requeue = true;
1795 __cap_set_timeouts(mdsc, ci);
1798 tried_invalidate = true;
1803 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1804 cap = rb_entry(p, struct ceph_cap, ci_node);
1807 /* avoid looping forever */
1808 if (mds >= cap->mds ||
1809 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1812 /* NOTE: no side-effects allowed, until we take s_mutex */
1815 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1816 cap_used &= ~ci->i_auth_cap->issued;
1818 revoking = cap->implemented & ~cap->issued;
1819 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1820 cap->mds, cap, ceph_cap_string(cap_used),
1821 ceph_cap_string(cap->issued),
1822 ceph_cap_string(cap->implemented),
1823 ceph_cap_string(revoking));
1825 if (cap == ci->i_auth_cap &&
1826 (cap->issued & CEPH_CAP_FILE_WR)) {
1827 /* request larger max_size from MDS? */
1828 if (ci->i_wanted_max_size > ci->i_max_size &&
1829 ci->i_wanted_max_size > ci->i_requested_max_size) {
1830 dout("requesting new max_size\n");
1834 /* approaching file_max? */
1835 if (__ceph_should_report_size(ci)) {
1836 dout("i_size approaching max_size\n");
1840 /* flush anything dirty? */
1841 if (cap == ci->i_auth_cap) {
1842 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1843 dout("flushing dirty caps\n");
1846 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1847 dout("flushing snap caps\n");
1852 /* completed revocation? going down and there are no caps? */
1853 if (revoking && (revoking & cap_used) == 0) {
1854 dout("completed revocation of %s\n",
1855 ceph_cap_string(cap->implemented & ~cap->issued));
1859 /* want more caps from mds? */
1860 if (want & ~(cap->mds_wanted | cap->issued))
1863 /* things we might delay */
1864 if ((cap->issued & ~retain) == 0 &&
1865 cap->mds_wanted == want)
1866 continue; /* nope, all good */
1872 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1873 time_before(jiffies, ci->i_hold_caps_max)) {
1874 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1875 ceph_cap_string(cap->issued),
1876 ceph_cap_string(cap->issued & retain),
1877 ceph_cap_string(cap->mds_wanted),
1878 ceph_cap_string(want));
1884 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1885 dout(" skipping %p I_NOFLUSH set\n", inode);
1889 if (session && session != cap->session) {
1890 dout("oops, wrong session %p mutex\n", session);
1891 mutex_unlock(&session->s_mutex);
1895 session = cap->session;
1896 if (mutex_trylock(&session->s_mutex) == 0) {
1897 dout("inverting session/ino locks on %p\n",
1899 spin_unlock(&ci->i_ceph_lock);
1900 if (took_snap_rwsem) {
1901 up_read(&mdsc->snap_rwsem);
1902 took_snap_rwsem = 0;
1904 mutex_lock(&session->s_mutex);
1909 /* kick flushing and flush snaps before sending normal
1911 if (cap == ci->i_auth_cap &&
1913 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
1914 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1915 __kick_flushing_caps(mdsc, session, ci, 0);
1916 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1918 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
1919 __ceph_flush_snaps(ci, session);
1924 /* take snap_rwsem after session mutex */
1925 if (!took_snap_rwsem) {
1926 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1927 dout("inverting snap/in locks on %p\n",
1929 spin_unlock(&ci->i_ceph_lock);
1930 down_read(&mdsc->snap_rwsem);
1931 took_snap_rwsem = 1;
1934 took_snap_rwsem = 1;
1937 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1938 flushing = __mark_caps_flushing(inode, session, false,
1944 spin_lock(&mdsc->cap_dirty_lock);
1945 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1946 spin_unlock(&mdsc->cap_dirty_lock);
1949 mds = cap->mds; /* remember mds, so we don't repeat */
1952 /* __send_cap drops i_ceph_lock */
1953 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, false,
1954 cap_used, want, retain, flushing,
1955 flush_tid, oldest_flush_tid);
1956 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1960 * Reschedule delayed caps release if we delayed anything,
1963 if (delayed && is_delayed)
1964 force_requeue = true; /* __send_cap delayed release; requeue */
1965 if (!delayed && !is_delayed)
1966 __cap_delay_cancel(mdsc, ci);
1967 else if (!is_delayed || force_requeue)
1968 __cap_delay_requeue(mdsc, ci);
1970 spin_unlock(&ci->i_ceph_lock);
1972 if (queue_invalidate)
1973 ceph_queue_invalidate(inode);
1976 mutex_unlock(&session->s_mutex);
1977 if (took_snap_rwsem)
1978 up_read(&mdsc->snap_rwsem);
1982 * Try to flush dirty caps back to the auth mds.
1984 static int try_flush_caps(struct inode *inode, u64 *ptid)
1986 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1987 struct ceph_inode_info *ci = ceph_inode(inode);
1988 struct ceph_mds_session *session = NULL;
1990 u64 flush_tid = 0, oldest_flush_tid = 0;
1993 spin_lock(&ci->i_ceph_lock);
1994 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1995 spin_unlock(&ci->i_ceph_lock);
1996 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1999 if (ci->i_dirty_caps && ci->i_auth_cap) {
2000 struct ceph_cap *cap = ci->i_auth_cap;
2001 int used = __ceph_caps_used(ci);
2002 int want = __ceph_caps_wanted(ci);
2005 if (!session || session != cap->session) {
2006 spin_unlock(&ci->i_ceph_lock);
2008 mutex_unlock(&session->s_mutex);
2009 session = cap->session;
2010 mutex_lock(&session->s_mutex);
2013 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2014 spin_unlock(&ci->i_ceph_lock);
2018 flushing = __mark_caps_flushing(inode, session, true,
2019 &flush_tid, &oldest_flush_tid);
2021 /* __send_cap drops i_ceph_lock */
2022 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, true,
2023 used, want, (cap->issued | cap->implemented),
2024 flushing, flush_tid, oldest_flush_tid);
2027 spin_lock(&ci->i_ceph_lock);
2028 __cap_delay_requeue(mdsc, ci);
2029 spin_unlock(&ci->i_ceph_lock);
2032 if (!list_empty(&ci->i_cap_flush_list)) {
2033 struct ceph_cap_flush *cf =
2034 list_last_entry(&ci->i_cap_flush_list,
2035 struct ceph_cap_flush, i_list);
2037 flush_tid = cf->tid;
2039 flushing = ci->i_flushing_caps;
2040 spin_unlock(&ci->i_ceph_lock);
2044 mutex_unlock(&session->s_mutex);
2051 * Return true if we've flushed caps through the given flush_tid.
2053 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2055 struct ceph_inode_info *ci = ceph_inode(inode);
2058 spin_lock(&ci->i_ceph_lock);
2059 if (!list_empty(&ci->i_cap_flush_list)) {
2060 struct ceph_cap_flush * cf =
2061 list_first_entry(&ci->i_cap_flush_list,
2062 struct ceph_cap_flush, i_list);
2063 if (cf->tid <= flush_tid)
2066 spin_unlock(&ci->i_ceph_lock);
2071 * wait for any unsafe requests to complete.
2073 static int unsafe_request_wait(struct inode *inode)
2075 struct ceph_inode_info *ci = ceph_inode(inode);
2076 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2079 spin_lock(&ci->i_unsafe_lock);
2080 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2081 req1 = list_last_entry(&ci->i_unsafe_dirops,
2082 struct ceph_mds_request,
2084 ceph_mdsc_get_request(req1);
2086 if (!list_empty(&ci->i_unsafe_iops)) {
2087 req2 = list_last_entry(&ci->i_unsafe_iops,
2088 struct ceph_mds_request,
2089 r_unsafe_target_item);
2090 ceph_mdsc_get_request(req2);
2092 spin_unlock(&ci->i_unsafe_lock);
2094 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2095 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2097 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2098 ceph_timeout_jiffies(req1->r_timeout));
2101 ceph_mdsc_put_request(req1);
2104 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2105 ceph_timeout_jiffies(req2->r_timeout));
2108 ceph_mdsc_put_request(req2);
2113 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2115 struct inode *inode = file->f_mapping->host;
2116 struct ceph_inode_info *ci = ceph_inode(inode);
2121 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2123 ret = file_write_and_wait_range(file, start, end);
2132 dirty = try_flush_caps(inode, &flush_tid);
2133 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2135 ret = unsafe_request_wait(inode);
2138 * only wait on non-file metadata writeback (the mds
2139 * can recover size and mtime, so we don't need to
2142 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2143 ret = wait_event_interruptible(ci->i_cap_wq,
2144 caps_are_flushed(inode, flush_tid));
2146 inode_unlock(inode);
2148 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2153 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2154 * queue inode for flush but don't do so immediately, because we can
2155 * get by with fewer MDS messages if we wait for data writeback to
2158 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2160 struct ceph_inode_info *ci = ceph_inode(inode);
2164 int wait = wbc->sync_mode == WB_SYNC_ALL;
2166 dout("write_inode %p wait=%d\n", inode, wait);
2168 dirty = try_flush_caps(inode, &flush_tid);
2170 err = wait_event_interruptible(ci->i_cap_wq,
2171 caps_are_flushed(inode, flush_tid));
2173 struct ceph_mds_client *mdsc =
2174 ceph_sb_to_client(inode->i_sb)->mdsc;
2176 spin_lock(&ci->i_ceph_lock);
2177 if (__ceph_caps_dirty(ci))
2178 __cap_delay_requeue_front(mdsc, ci);
2179 spin_unlock(&ci->i_ceph_lock);
2184 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2185 struct ceph_mds_session *session,
2186 struct ceph_inode_info *ci,
2187 u64 oldest_flush_tid)
2188 __releases(ci->i_ceph_lock)
2189 __acquires(ci->i_ceph_lock)
2191 struct inode *inode = &ci->vfs_inode;
2192 struct ceph_cap *cap;
2193 struct ceph_cap_flush *cf;
2197 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2198 if (cf->tid < first_tid)
2201 cap = ci->i_auth_cap;
2202 if (!(cap && cap->session == session)) {
2203 pr_err("%p auth cap %p not mds%d ???\n",
2204 inode, cap, session->s_mds);
2208 first_tid = cf->tid + 1;
2211 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2212 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2213 ci->i_ceph_flags |= CEPH_I_NODELAY;
2214 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2215 false, __ceph_caps_used(ci),
2216 __ceph_caps_wanted(ci),
2217 cap->issued | cap->implemented,
2218 cf->caps, cf->tid, oldest_flush_tid);
2220 pr_err("kick_flushing_caps: error sending "
2221 "cap flush, ino (%llx.%llx) "
2222 "tid %llu flushing %s\n",
2223 ceph_vinop(inode), cf->tid,
2224 ceph_cap_string(cf->caps));
2227 struct ceph_cap_snap *capsnap =
2228 container_of(cf, struct ceph_cap_snap,
2230 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2231 inode, capsnap, cf->tid,
2232 ceph_cap_string(capsnap->dirty));
2234 refcount_inc(&capsnap->nref);
2235 spin_unlock(&ci->i_ceph_lock);
2237 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2240 pr_err("kick_flushing_caps: error sending "
2241 "cap flushsnap, ino (%llx.%llx) "
2242 "tid %llu follows %llu\n",
2243 ceph_vinop(inode), cf->tid,
2247 ceph_put_cap_snap(capsnap);
2250 spin_lock(&ci->i_ceph_lock);
2254 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2255 struct ceph_mds_session *session)
2257 struct ceph_inode_info *ci;
2258 struct ceph_cap *cap;
2259 u64 oldest_flush_tid;
2261 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2263 spin_lock(&mdsc->cap_dirty_lock);
2264 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2265 spin_unlock(&mdsc->cap_dirty_lock);
2267 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2268 spin_lock(&ci->i_ceph_lock);
2269 cap = ci->i_auth_cap;
2270 if (!(cap && cap->session == session)) {
2271 pr_err("%p auth cap %p not mds%d ???\n",
2272 &ci->vfs_inode, cap, session->s_mds);
2273 spin_unlock(&ci->i_ceph_lock);
2279 * if flushing caps were revoked, we re-send the cap flush
2280 * in client reconnect stage. This guarantees MDS * processes
2281 * the cap flush message before issuing the flushing caps to
2284 if ((cap->issued & ci->i_flushing_caps) !=
2285 ci->i_flushing_caps) {
2286 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2287 __kick_flushing_caps(mdsc, session, ci,
2290 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2293 spin_unlock(&ci->i_ceph_lock);
2297 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2298 struct ceph_mds_session *session)
2300 struct ceph_inode_info *ci;
2301 struct ceph_cap *cap;
2302 u64 oldest_flush_tid;
2304 dout("kick_flushing_caps mds%d\n", session->s_mds);
2306 spin_lock(&mdsc->cap_dirty_lock);
2307 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2308 spin_unlock(&mdsc->cap_dirty_lock);
2310 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2311 spin_lock(&ci->i_ceph_lock);
2312 cap = ci->i_auth_cap;
2313 if (!(cap && cap->session == session)) {
2314 pr_err("%p auth cap %p not mds%d ???\n",
2315 &ci->vfs_inode, cap, session->s_mds);
2316 spin_unlock(&ci->i_ceph_lock);
2319 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2320 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2321 __kick_flushing_caps(mdsc, session, ci,
2324 spin_unlock(&ci->i_ceph_lock);
2328 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2329 struct ceph_mds_session *session,
2330 struct inode *inode)
2331 __releases(ci->i_ceph_lock)
2333 struct ceph_inode_info *ci = ceph_inode(inode);
2334 struct ceph_cap *cap;
2336 cap = ci->i_auth_cap;
2337 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2338 ceph_cap_string(ci->i_flushing_caps));
2340 if (!list_empty(&ci->i_cap_flush_list)) {
2341 u64 oldest_flush_tid;
2342 spin_lock(&mdsc->cap_dirty_lock);
2343 list_move_tail(&ci->i_flushing_item,
2344 &cap->session->s_cap_flushing);
2345 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2346 spin_unlock(&mdsc->cap_dirty_lock);
2348 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2349 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2350 spin_unlock(&ci->i_ceph_lock);
2352 spin_unlock(&ci->i_ceph_lock);
2358 * Take references to capabilities we hold, so that we don't release
2359 * them to the MDS prematurely.
2361 * Protected by i_ceph_lock.
2363 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2364 bool snap_rwsem_locked)
2366 if (got & CEPH_CAP_PIN)
2368 if (got & CEPH_CAP_FILE_RD)
2370 if (got & CEPH_CAP_FILE_CACHE)
2371 ci->i_rdcache_ref++;
2372 if (got & CEPH_CAP_FILE_WR) {
2373 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2374 BUG_ON(!snap_rwsem_locked);
2375 ci->i_head_snapc = ceph_get_snap_context(
2376 ci->i_snap_realm->cached_context);
2380 if (got & CEPH_CAP_FILE_BUFFER) {
2381 if (ci->i_wb_ref == 0)
2382 ihold(&ci->vfs_inode);
2384 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2385 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2390 * Try to grab cap references. Specify those refs we @want, and the
2391 * minimal set we @need. Also include the larger offset we are writing
2392 * to (when applicable), and check against max_size here as well.
2393 * Note that caller is responsible for ensuring max_size increases are
2394 * requested from the MDS.
2396 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2397 loff_t endoff, bool nonblock, int *got, int *err)
2399 struct inode *inode = &ci->vfs_inode;
2400 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2402 int have, implemented;
2404 bool snap_rwsem_locked = false;
2406 dout("get_cap_refs %p need %s want %s\n", inode,
2407 ceph_cap_string(need), ceph_cap_string(want));
2410 spin_lock(&ci->i_ceph_lock);
2412 /* make sure file is actually open */
2413 file_wanted = __ceph_caps_file_wanted(ci);
2414 if ((file_wanted & need) != need) {
2415 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2416 ceph_cap_string(need), ceph_cap_string(file_wanted));
2422 /* finish pending truncate */
2423 while (ci->i_truncate_pending) {
2424 spin_unlock(&ci->i_ceph_lock);
2425 if (snap_rwsem_locked) {
2426 up_read(&mdsc->snap_rwsem);
2427 snap_rwsem_locked = false;
2429 __ceph_do_pending_vmtruncate(inode);
2430 spin_lock(&ci->i_ceph_lock);
2433 have = __ceph_caps_issued(ci, &implemented);
2435 if (have & need & CEPH_CAP_FILE_WR) {
2436 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2437 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2438 inode, endoff, ci->i_max_size);
2439 if (endoff > ci->i_requested_max_size) {
2446 * If a sync write is in progress, we must wait, so that we
2447 * can get a final snapshot value for size+mtime.
2449 if (__ceph_have_pending_cap_snap(ci)) {
2450 dout("get_cap_refs %p cap_snap_pending\n", inode);
2455 if ((have & need) == need) {
2457 * Look at (implemented & ~have & not) so that we keep waiting
2458 * on transition from wanted -> needed caps. This is needed
2459 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2460 * going before a prior buffered writeback happens.
2462 int not = want & ~(have & need);
2463 int revoking = implemented & ~have;
2464 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2465 inode, ceph_cap_string(have), ceph_cap_string(not),
2466 ceph_cap_string(revoking));
2467 if ((revoking & not) == 0) {
2468 if (!snap_rwsem_locked &&
2469 !ci->i_head_snapc &&
2470 (need & CEPH_CAP_FILE_WR)) {
2471 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2473 * we can not call down_read() when
2474 * task isn't in TASK_RUNNING state
2482 spin_unlock(&ci->i_ceph_lock);
2483 down_read(&mdsc->snap_rwsem);
2484 snap_rwsem_locked = true;
2487 snap_rwsem_locked = true;
2489 *got = need | (have & want);
2490 if ((need & CEPH_CAP_FILE_RD) &&
2491 !(*got & CEPH_CAP_FILE_CACHE))
2492 ceph_disable_fscache_readpage(ci);
2493 __take_cap_refs(ci, *got, true);
2497 int session_readonly = false;
2498 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2499 struct ceph_mds_session *s = ci->i_auth_cap->session;
2500 spin_lock(&s->s_cap_lock);
2501 session_readonly = s->s_readonly;
2502 spin_unlock(&s->s_cap_lock);
2504 if (session_readonly) {
2505 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2506 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2512 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2514 if (READ_ONCE(mdsc->fsc->mount_state) ==
2515 CEPH_MOUNT_SHUTDOWN) {
2516 dout("get_cap_refs %p forced umount\n", inode);
2521 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2522 if (need & ~(mds_wanted & need)) {
2523 dout("get_cap_refs %p caps were dropped"
2524 " (session killed?)\n", inode);
2529 if (!(file_wanted & ~mds_wanted))
2530 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2533 dout("get_cap_refs %p have %s needed %s\n", inode,
2534 ceph_cap_string(have), ceph_cap_string(need));
2537 spin_unlock(&ci->i_ceph_lock);
2538 if (snap_rwsem_locked)
2539 up_read(&mdsc->snap_rwsem);
2541 dout("get_cap_refs %p ret %d got %s\n", inode,
2542 ret, ceph_cap_string(*got));
2547 * Check the offset we are writing up to against our current
2548 * max_size. If necessary, tell the MDS we want to write to
2551 static void check_max_size(struct inode *inode, loff_t endoff)
2553 struct ceph_inode_info *ci = ceph_inode(inode);
2556 /* do we need to explicitly request a larger max_size? */
2557 spin_lock(&ci->i_ceph_lock);
2558 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2559 dout("write %p at large endoff %llu, req max_size\n",
2561 ci->i_wanted_max_size = endoff;
2563 /* duplicate ceph_check_caps()'s logic */
2564 if (ci->i_auth_cap &&
2565 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2566 ci->i_wanted_max_size > ci->i_max_size &&
2567 ci->i_wanted_max_size > ci->i_requested_max_size)
2569 spin_unlock(&ci->i_ceph_lock);
2571 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2574 int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want, int *got)
2578 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2579 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
2580 ret = ceph_pool_perm_check(ci, need);
2584 ret = try_get_cap_refs(ci, need, want, 0, true, got, &err);
2586 if (err == -EAGAIN) {
2588 } else if (err < 0) {
2596 * Wait for caps, and take cap references. If we can't get a WR cap
2597 * due to a small max_size, make sure we check_max_size (and possibly
2598 * ask the mds) so we don't get hung up indefinitely.
2600 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2601 loff_t endoff, int *got, struct page **pinned_page)
2603 int _got, ret, err = 0;
2605 ret = ceph_pool_perm_check(ci, need);
2611 check_max_size(&ci->vfs_inode, endoff);
2615 ret = try_get_cap_refs(ci, need, want, endoff,
2616 false, &_got, &err);
2623 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2624 add_wait_queue(&ci->i_cap_wq, &wait);
2626 while (!try_get_cap_refs(ci, need, want, endoff,
2627 true, &_got, &err)) {
2628 if (signal_pending(current)) {
2632 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2635 remove_wait_queue(&ci->i_cap_wq, &wait);
2643 if (err == -ESTALE) {
2644 /* session was killed, try renew caps */
2645 ret = ceph_renew_caps(&ci->vfs_inode);
2652 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2653 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2654 i_size_read(&ci->vfs_inode) > 0) {
2656 find_get_page(ci->vfs_inode.i_mapping, 0);
2658 if (PageUptodate(page)) {
2659 *pinned_page = page;
2665 * drop cap refs first because getattr while
2666 * holding * caps refs can cause deadlock.
2668 ceph_put_cap_refs(ci, _got);
2672 * getattr request will bring inline data into
2675 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2676 CEPH_STAT_CAP_INLINE_DATA,
2685 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2686 ceph_fscache_revalidate_cookie(ci);
2693 * Take cap refs. Caller must already know we hold at least one ref
2694 * on the caps in question or we don't know this is safe.
2696 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2698 spin_lock(&ci->i_ceph_lock);
2699 __take_cap_refs(ci, caps, false);
2700 spin_unlock(&ci->i_ceph_lock);
2705 * drop cap_snap that is not associated with any snapshot.
2706 * we don't need to send FLUSHSNAP message for it.
2708 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2709 struct ceph_cap_snap *capsnap)
2711 if (!capsnap->need_flush &&
2712 !capsnap->writing && !capsnap->dirty_pages) {
2713 dout("dropping cap_snap %p follows %llu\n",
2714 capsnap, capsnap->follows);
2715 BUG_ON(capsnap->cap_flush.tid > 0);
2716 ceph_put_snap_context(capsnap->context);
2717 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2718 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2720 list_del(&capsnap->ci_item);
2721 ceph_put_cap_snap(capsnap);
2730 * If we released the last ref on any given cap, call ceph_check_caps
2731 * to release (or schedule a release).
2733 * If we are releasing a WR cap (from a sync write), finalize any affected
2734 * cap_snap, and wake up any waiters.
2736 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2738 struct inode *inode = &ci->vfs_inode;
2739 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2741 spin_lock(&ci->i_ceph_lock);
2742 if (had & CEPH_CAP_PIN)
2744 if (had & CEPH_CAP_FILE_RD)
2745 if (--ci->i_rd_ref == 0)
2747 if (had & CEPH_CAP_FILE_CACHE)
2748 if (--ci->i_rdcache_ref == 0)
2750 if (had & CEPH_CAP_FILE_BUFFER) {
2751 if (--ci->i_wb_ref == 0) {
2755 dout("put_cap_refs %p wb %d -> %d (?)\n",
2756 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2758 if (had & CEPH_CAP_FILE_WR)
2759 if (--ci->i_wr_ref == 0) {
2761 if (__ceph_have_pending_cap_snap(ci)) {
2762 struct ceph_cap_snap *capsnap =
2763 list_last_entry(&ci->i_cap_snaps,
2764 struct ceph_cap_snap,
2766 capsnap->writing = 0;
2767 if (ceph_try_drop_cap_snap(ci, capsnap))
2769 else if (__ceph_finish_cap_snap(ci, capsnap))
2773 if (ci->i_wrbuffer_ref_head == 0 &&
2774 ci->i_dirty_caps == 0 &&
2775 ci->i_flushing_caps == 0) {
2776 BUG_ON(!ci->i_head_snapc);
2777 ceph_put_snap_context(ci->i_head_snapc);
2778 ci->i_head_snapc = NULL;
2780 /* see comment in __ceph_remove_cap() */
2781 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2782 drop_inode_snap_realm(ci);
2784 spin_unlock(&ci->i_ceph_lock);
2786 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2787 last ? " last" : "", put ? " put" : "");
2789 if (last && !flushsnaps)
2790 ceph_check_caps(ci, 0, NULL);
2791 else if (flushsnaps)
2792 ceph_flush_snaps(ci, NULL);
2794 wake_up_all(&ci->i_cap_wq);
2800 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2801 * context. Adjust per-snap dirty page accounting as appropriate.
2802 * Once all dirty data for a cap_snap is flushed, flush snapped file
2803 * metadata back to the MDS. If we dropped the last ref, call
2806 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2807 struct ceph_snap_context *snapc)
2809 struct inode *inode = &ci->vfs_inode;
2810 struct ceph_cap_snap *capsnap = NULL;
2814 bool flush_snaps = false;
2815 bool complete_capsnap = false;
2817 spin_lock(&ci->i_ceph_lock);
2818 ci->i_wrbuffer_ref -= nr;
2819 if (ci->i_wrbuffer_ref == 0) {
2824 if (ci->i_head_snapc == snapc) {
2825 ci->i_wrbuffer_ref_head -= nr;
2826 if (ci->i_wrbuffer_ref_head == 0 &&
2827 ci->i_wr_ref == 0 &&
2828 ci->i_dirty_caps == 0 &&
2829 ci->i_flushing_caps == 0) {
2830 BUG_ON(!ci->i_head_snapc);
2831 ceph_put_snap_context(ci->i_head_snapc);
2832 ci->i_head_snapc = NULL;
2834 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2836 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2837 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2838 last ? " LAST" : "");
2840 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2841 if (capsnap->context == snapc) {
2847 capsnap->dirty_pages -= nr;
2848 if (capsnap->dirty_pages == 0) {
2849 complete_capsnap = true;
2850 if (!capsnap->writing) {
2851 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2854 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2859 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2860 " snap %lld %d/%d -> %d/%d %s%s\n",
2861 inode, capsnap, capsnap->context->seq,
2862 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2863 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2864 last ? " (wrbuffer last)" : "",
2865 complete_capsnap ? " (complete capsnap)" : "");
2868 spin_unlock(&ci->i_ceph_lock);
2871 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2872 } else if (flush_snaps) {
2873 ceph_flush_snaps(ci, NULL);
2875 if (complete_capsnap)
2876 wake_up_all(&ci->i_cap_wq);
2882 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2884 static void invalidate_aliases(struct inode *inode)
2886 struct dentry *dn, *prev = NULL;
2888 dout("invalidate_aliases inode %p\n", inode);
2889 d_prune_aliases(inode);
2891 * For non-directory inode, d_find_alias() only returns
2892 * hashed dentry. After calling d_invalidate(), the
2893 * dentry becomes unhashed.
2895 * For directory inode, d_find_alias() can return
2896 * unhashed dentry. But directory inode should have
2897 * one alias at most.
2899 while ((dn = d_find_alias(inode))) {
2914 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2915 * actually be a revocation if it specifies a smaller cap set.)
2917 * caller holds s_mutex and i_ceph_lock, we drop both.
2919 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2920 struct inode *inode, struct ceph_mds_caps *grant,
2921 struct ceph_string **pns, u64 inline_version,
2922 void *inline_data, u32 inline_len,
2923 struct ceph_buffer *xattr_buf,
2924 struct ceph_mds_session *session,
2925 struct ceph_cap *cap, int issued)
2926 __releases(ci->i_ceph_lock)
2927 __releases(mdsc->snap_rwsem)
2929 struct ceph_inode_info *ci = ceph_inode(inode);
2930 int mds = session->s_mds;
2931 int seq = le32_to_cpu(grant->seq);
2932 int newcaps = le32_to_cpu(grant->caps);
2933 int used, wanted, dirty;
2934 u64 size = le64_to_cpu(grant->size);
2935 u64 max_size = le64_to_cpu(grant->max_size);
2936 struct timespec mtime, atime, ctime;
2939 bool writeback = false;
2940 bool queue_trunc = false;
2941 bool queue_invalidate = false;
2942 bool deleted_inode = false;
2943 bool fill_inline = false;
2945 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2946 inode, cap, mds, seq, ceph_cap_string(newcaps));
2947 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2952 * auth mds of the inode changed. we received the cap export message,
2953 * but still haven't received the cap import message. handle_cap_export
2954 * updated the new auth MDS' cap.
2956 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2957 * that was sent before the cap import message. So don't remove caps.
2959 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2960 WARN_ON(cap != ci->i_auth_cap);
2961 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2963 newcaps |= cap->issued;
2967 * If CACHE is being revoked, and we have no dirty buffers,
2968 * try to invalidate (once). (If there are dirty buffers, we
2969 * will invalidate _after_ writeback.)
2971 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2972 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2973 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2974 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2975 if (try_nonblocking_invalidate(inode)) {
2976 /* there were locked pages.. invalidate later
2977 in a separate thread. */
2978 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2979 queue_invalidate = true;
2980 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2985 /* side effects now are allowed */
2986 cap->cap_gen = session->s_cap_gen;
2989 __check_cap_issue(ci, cap, newcaps);
2991 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2992 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2993 inode->i_mode = le32_to_cpu(grant->mode);
2994 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2995 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2996 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2997 from_kuid(&init_user_ns, inode->i_uid),
2998 from_kgid(&init_user_ns, inode->i_gid));
3001 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3002 (issued & CEPH_CAP_LINK_EXCL) == 0) {
3003 set_nlink(inode, le32_to_cpu(grant->nlink));
3004 if (inode->i_nlink == 0 &&
3005 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3006 deleted_inode = true;
3009 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
3010 int len = le32_to_cpu(grant->xattr_len);
3011 u64 version = le64_to_cpu(grant->xattr_version);
3013 if (version > ci->i_xattrs.version) {
3014 dout(" got new xattrs v%llu on %p len %d\n",
3015 version, inode, len);
3016 if (ci->i_xattrs.blob)
3017 ceph_buffer_put(ci->i_xattrs.blob);
3018 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3019 ci->i_xattrs.version = version;
3020 ceph_forget_all_cached_acls(inode);
3024 if (newcaps & CEPH_CAP_ANY_RD) {
3025 /* ctime/mtime/atime? */
3026 ceph_decode_timespec(&mtime, &grant->mtime);
3027 ceph_decode_timespec(&atime, &grant->atime);
3028 ceph_decode_timespec(&ctime, &grant->ctime);
3029 ceph_fill_file_time(inode, issued,
3030 le32_to_cpu(grant->time_warp_seq),
3031 &ctime, &mtime, &atime);
3034 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3035 /* file layout may have changed */
3036 s64 old_pool = ci->i_layout.pool_id;
3037 struct ceph_string *old_ns;
3039 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3040 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3041 lockdep_is_held(&ci->i_ceph_lock));
3042 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
3044 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
3045 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3049 /* size/truncate_seq? */
3050 queue_trunc = ceph_fill_file_size(inode, issued,
3051 le32_to_cpu(grant->truncate_seq),
3052 le64_to_cpu(grant->truncate_size),
3056 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3057 if (max_size != ci->i_max_size) {
3058 dout("max_size %lld -> %llu\n",
3059 ci->i_max_size, max_size);
3060 ci->i_max_size = max_size;
3061 if (max_size >= ci->i_wanted_max_size) {
3062 ci->i_wanted_max_size = 0; /* reset */
3063 ci->i_requested_max_size = 0;
3066 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3067 ci->i_wanted_max_size > ci->i_requested_max_size) {
3068 /* CEPH_CAP_OP_IMPORT */
3073 /* check cap bits */
3074 wanted = __ceph_caps_wanted(ci);
3075 used = __ceph_caps_used(ci);
3076 dirty = __ceph_caps_dirty(ci);
3077 dout(" my wanted = %s, used = %s, dirty %s\n",
3078 ceph_cap_string(wanted),
3079 ceph_cap_string(used),
3080 ceph_cap_string(dirty));
3081 if (wanted != le32_to_cpu(grant->wanted)) {
3082 dout("mds wanted %s -> %s\n",
3083 ceph_cap_string(le32_to_cpu(grant->wanted)),
3084 ceph_cap_string(wanted));
3085 /* imported cap may not have correct mds_wanted */
3086 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
3090 /* revocation, grant, or no-op? */
3091 if (cap->issued & ~newcaps) {
3092 int revoking = cap->issued & ~newcaps;
3094 dout("revocation: %s -> %s (revoking %s)\n",
3095 ceph_cap_string(cap->issued),
3096 ceph_cap_string(newcaps),
3097 ceph_cap_string(revoking));
3098 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3099 writeback = true; /* initiate writeback; will delay ack */
3100 else if (revoking == CEPH_CAP_FILE_CACHE &&
3101 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3103 ; /* do nothing yet, invalidation will be queued */
3104 else if (cap == ci->i_auth_cap)
3105 check_caps = 1; /* check auth cap only */
3107 check_caps = 2; /* check all caps */
3108 cap->issued = newcaps;
3109 cap->implemented |= newcaps;
3110 } else if (cap->issued == newcaps) {
3111 dout("caps unchanged: %s -> %s\n",
3112 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3114 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3115 ceph_cap_string(newcaps));
3116 /* non-auth MDS is revoking the newly grant caps ? */
3117 if (cap == ci->i_auth_cap &&
3118 __ceph_caps_revoking_other(ci, cap, newcaps))
3121 cap->issued = newcaps;
3122 cap->implemented |= newcaps; /* add bits only, to
3123 * avoid stepping on a
3124 * pending revocation */
3127 BUG_ON(cap->issued & ~cap->implemented);
3129 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
3130 ci->i_inline_version = inline_version;
3131 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3132 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3136 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3137 if (newcaps & ~issued)
3139 kick_flushing_inode_caps(mdsc, session, inode);
3140 up_read(&mdsc->snap_rwsem);
3142 spin_unlock(&ci->i_ceph_lock);
3146 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
3149 ceph_queue_vmtruncate(inode);
3153 * queue inode for writeback: we can't actually call
3154 * filemap_write_and_wait, etc. from message handler
3157 ceph_queue_writeback(inode);
3158 if (queue_invalidate)
3159 ceph_queue_invalidate(inode);
3161 invalidate_aliases(inode);
3163 wake_up_all(&ci->i_cap_wq);
3165 if (check_caps == 1)
3166 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3168 else if (check_caps == 2)
3169 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3171 mutex_unlock(&session->s_mutex);
3175 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3176 * MDS has been safely committed.
3178 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3179 struct ceph_mds_caps *m,
3180 struct ceph_mds_session *session,
3181 struct ceph_cap *cap)
3182 __releases(ci->i_ceph_lock)
3184 struct ceph_inode_info *ci = ceph_inode(inode);
3185 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3186 struct ceph_cap_flush *cf, *tmp_cf;
3187 LIST_HEAD(to_remove);
3188 unsigned seq = le32_to_cpu(m->seq);
3189 int dirty = le32_to_cpu(m->dirty);
3195 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3196 if (cf->tid == flush_tid)
3198 if (cf->caps == 0) /* capsnap */
3200 if (cf->tid <= flush_tid) {
3201 if (__finish_cap_flush(NULL, ci, cf))
3203 list_add_tail(&cf->i_list, &to_remove);
3205 cleaned &= ~cf->caps;
3211 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3212 " flushing %s -> %s\n",
3213 inode, session->s_mds, seq, ceph_cap_string(dirty),
3214 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3215 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3217 if (list_empty(&to_remove) && !cleaned)
3220 ci->i_flushing_caps &= ~cleaned;
3222 spin_lock(&mdsc->cap_dirty_lock);
3224 list_for_each_entry(cf, &to_remove, i_list) {
3225 if (__finish_cap_flush(mdsc, NULL, cf))
3229 if (ci->i_flushing_caps == 0) {
3230 if (list_empty(&ci->i_cap_flush_list)) {
3231 list_del_init(&ci->i_flushing_item);
3232 if (!list_empty(&session->s_cap_flushing)) {
3233 dout(" mds%d still flushing cap on %p\n",
3235 &list_first_entry(&session->s_cap_flushing,
3236 struct ceph_inode_info,
3237 i_flushing_item)->vfs_inode);
3240 mdsc->num_cap_flushing--;
3241 dout(" inode %p now !flushing\n", inode);
3243 if (ci->i_dirty_caps == 0) {
3244 dout(" inode %p now clean\n", inode);
3245 BUG_ON(!list_empty(&ci->i_dirty_item));
3247 if (ci->i_wr_ref == 0 &&
3248 ci->i_wrbuffer_ref_head == 0) {
3249 BUG_ON(!ci->i_head_snapc);
3250 ceph_put_snap_context(ci->i_head_snapc);
3251 ci->i_head_snapc = NULL;
3254 BUG_ON(list_empty(&ci->i_dirty_item));
3257 spin_unlock(&mdsc->cap_dirty_lock);
3260 spin_unlock(&ci->i_ceph_lock);
3262 while (!list_empty(&to_remove)) {
3263 cf = list_first_entry(&to_remove,
3264 struct ceph_cap_flush, i_list);
3265 list_del(&cf->i_list);
3266 ceph_free_cap_flush(cf);
3270 wake_up_all(&ci->i_cap_wq);
3272 wake_up_all(&mdsc->cap_flushing_wq);
3278 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3279 * throw away our cap_snap.
3281 * Caller hold s_mutex.
3283 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3284 struct ceph_mds_caps *m,
3285 struct ceph_mds_session *session)
3287 struct ceph_inode_info *ci = ceph_inode(inode);
3288 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3289 u64 follows = le64_to_cpu(m->snap_follows);
3290 struct ceph_cap_snap *capsnap;
3291 bool flushed = false;
3292 bool wake_ci = false;
3293 bool wake_mdsc = false;
3295 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3296 inode, ci, session->s_mds, follows);
3298 spin_lock(&ci->i_ceph_lock);
3299 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3300 if (capsnap->follows == follows) {
3301 if (capsnap->cap_flush.tid != flush_tid) {
3302 dout(" cap_snap %p follows %lld tid %lld !="
3303 " %lld\n", capsnap, follows,
3304 flush_tid, capsnap->cap_flush.tid);
3310 dout(" skipping cap_snap %p follows %lld\n",
3311 capsnap, capsnap->follows);
3315 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3316 dout(" removing %p cap_snap %p follows %lld\n",
3317 inode, capsnap, follows);
3318 list_del(&capsnap->ci_item);
3319 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3322 spin_lock(&mdsc->cap_dirty_lock);
3324 if (list_empty(&ci->i_cap_flush_list))
3325 list_del_init(&ci->i_flushing_item);
3327 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3330 spin_unlock(&mdsc->cap_dirty_lock);
3332 spin_unlock(&ci->i_ceph_lock);
3334 ceph_put_snap_context(capsnap->context);
3335 ceph_put_cap_snap(capsnap);
3337 wake_up_all(&ci->i_cap_wq);
3339 wake_up_all(&mdsc->cap_flushing_wq);
3345 * Handle TRUNC from MDS, indicating file truncation.
3347 * caller hold s_mutex.
3349 static void handle_cap_trunc(struct inode *inode,
3350 struct ceph_mds_caps *trunc,
3351 struct ceph_mds_session *session)
3352 __releases(ci->i_ceph_lock)
3354 struct ceph_inode_info *ci = ceph_inode(inode);
3355 int mds = session->s_mds;
3356 int seq = le32_to_cpu(trunc->seq);
3357 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3358 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3359 u64 size = le64_to_cpu(trunc->size);
3360 int implemented = 0;
3361 int dirty = __ceph_caps_dirty(ci);
3362 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3363 int queue_trunc = 0;
3365 issued |= implemented | dirty;
3367 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3368 inode, mds, seq, truncate_size, truncate_seq);
3369 queue_trunc = ceph_fill_file_size(inode, issued,
3370 truncate_seq, truncate_size, size);
3371 spin_unlock(&ci->i_ceph_lock);
3374 ceph_queue_vmtruncate(inode);
3378 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3379 * different one. If we are the most recent migration we've seen (as
3380 * indicated by mseq), make note of the migrating cap bits for the
3381 * duration (until we see the corresponding IMPORT).
3383 * caller holds s_mutex
3385 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3386 struct ceph_mds_cap_peer *ph,
3387 struct ceph_mds_session *session)
3389 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3390 struct ceph_mds_session *tsession = NULL;
3391 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3392 struct ceph_inode_info *ci = ceph_inode(inode);
3394 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3395 unsigned t_seq, t_mseq;
3397 int mds = session->s_mds;
3400 t_cap_id = le64_to_cpu(ph->cap_id);
3401 t_seq = le32_to_cpu(ph->seq);
3402 t_mseq = le32_to_cpu(ph->mseq);
3403 target = le32_to_cpu(ph->mds);
3405 t_cap_id = t_seq = t_mseq = 0;
3409 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3410 inode, ci, mds, mseq, target);
3412 spin_lock(&ci->i_ceph_lock);
3413 cap = __get_cap_for_mds(ci, mds);
3414 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3418 __ceph_remove_cap(cap, false);
3419 if (!ci->i_auth_cap)
3420 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3425 * now we know we haven't received the cap import message yet
3426 * because the exported cap still exist.
3429 issued = cap->issued;
3430 WARN_ON(issued != cap->implemented);
3432 tcap = __get_cap_for_mds(ci, target);
3434 /* already have caps from the target */
3435 if (tcap->cap_id == t_cap_id &&
3436 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3437 dout(" updating import cap %p mds%d\n", tcap, target);
3438 tcap->cap_id = t_cap_id;
3439 tcap->seq = t_seq - 1;
3440 tcap->issue_seq = t_seq - 1;
3441 tcap->mseq = t_mseq;
3442 tcap->issued |= issued;
3443 tcap->implemented |= issued;
3444 if (cap == ci->i_auth_cap)
3445 ci->i_auth_cap = tcap;
3447 if (!list_empty(&ci->i_cap_flush_list) &&
3448 ci->i_auth_cap == tcap) {
3449 spin_lock(&mdsc->cap_dirty_lock);
3450 list_move_tail(&ci->i_flushing_item,
3451 &tcap->session->s_cap_flushing);
3452 spin_unlock(&mdsc->cap_dirty_lock);
3455 __ceph_remove_cap(cap, false);
3457 } else if (tsession) {
3458 /* add placeholder for the export tagert */
3459 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3461 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3462 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3464 if (!list_empty(&ci->i_cap_flush_list) &&
3465 ci->i_auth_cap == tcap) {
3466 spin_lock(&mdsc->cap_dirty_lock);
3467 list_move_tail(&ci->i_flushing_item,
3468 &tcap->session->s_cap_flushing);
3469 spin_unlock(&mdsc->cap_dirty_lock);
3472 __ceph_remove_cap(cap, false);
3476 spin_unlock(&ci->i_ceph_lock);
3477 mutex_unlock(&session->s_mutex);
3479 /* open target session */
3480 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3481 if (!IS_ERR(tsession)) {
3483 mutex_lock(&session->s_mutex);
3484 mutex_lock_nested(&tsession->s_mutex,
3485 SINGLE_DEPTH_NESTING);
3487 mutex_lock(&tsession->s_mutex);
3488 mutex_lock_nested(&session->s_mutex,
3489 SINGLE_DEPTH_NESTING);
3491 new_cap = ceph_get_cap(mdsc, NULL);
3500 spin_unlock(&ci->i_ceph_lock);
3501 mutex_unlock(&session->s_mutex);
3503 mutex_unlock(&tsession->s_mutex);
3504 ceph_put_mds_session(tsession);
3507 ceph_put_cap(mdsc, new_cap);
3511 * Handle cap IMPORT.
3513 * caller holds s_mutex. acquires i_ceph_lock
3515 static void handle_cap_import(struct ceph_mds_client *mdsc,
3516 struct inode *inode, struct ceph_mds_caps *im,
3517 struct ceph_mds_cap_peer *ph,
3518 struct ceph_mds_session *session,
3519 struct ceph_cap **target_cap, int *old_issued)
3520 __acquires(ci->i_ceph_lock)
3522 struct ceph_inode_info *ci = ceph_inode(inode);
3523 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3524 int mds = session->s_mds;
3526 unsigned caps = le32_to_cpu(im->caps);
3527 unsigned wanted = le32_to_cpu(im->wanted);
3528 unsigned seq = le32_to_cpu(im->seq);
3529 unsigned mseq = le32_to_cpu(im->migrate_seq);
3530 u64 realmino = le64_to_cpu(im->realm);
3531 u64 cap_id = le64_to_cpu(im->cap_id);
3536 p_cap_id = le64_to_cpu(ph->cap_id);
3537 peer = le32_to_cpu(ph->mds);
3543 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3544 inode, ci, mds, mseq, peer);
3547 spin_lock(&ci->i_ceph_lock);
3548 cap = __get_cap_for_mds(ci, mds);
3551 spin_unlock(&ci->i_ceph_lock);
3552 new_cap = ceph_get_cap(mdsc, NULL);
3558 ceph_put_cap(mdsc, new_cap);
3563 __ceph_caps_issued(ci, &issued);
3564 issued |= __ceph_caps_dirty(ci);
3566 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3567 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3569 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3570 if (ocap && ocap->cap_id == p_cap_id) {
3571 dout(" remove export cap %p mds%d flags %d\n",
3572 ocap, peer, ph->flags);
3573 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3574 (ocap->seq != le32_to_cpu(ph->seq) ||
3575 ocap->mseq != le32_to_cpu(ph->mseq))) {
3576 pr_err("handle_cap_import: mismatched seq/mseq: "
3577 "ino (%llx.%llx) mds%d seq %d mseq %d "
3578 "importer mds%d has peer seq %d mseq %d\n",
3579 ceph_vinop(inode), peer, ocap->seq,
3580 ocap->mseq, mds, le32_to_cpu(ph->seq),
3581 le32_to_cpu(ph->mseq));
3583 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3586 /* make sure we re-request max_size, if necessary */
3587 ci->i_requested_max_size = 0;
3589 *old_issued = issued;
3594 * Handle a caps message from the MDS.
3596 * Identify the appropriate session, inode, and call the right handler
3597 * based on the cap op.
3599 void ceph_handle_caps(struct ceph_mds_session *session,
3600 struct ceph_msg *msg)
3602 struct ceph_mds_client *mdsc = session->s_mdsc;
3603 struct super_block *sb = mdsc->fsc->sb;
3604 struct inode *inode;
3605 struct ceph_inode_info *ci;
3606 struct ceph_cap *cap;
3607 struct ceph_mds_caps *h;
3608 struct ceph_mds_cap_peer *peer = NULL;
3609 struct ceph_snap_realm *realm = NULL;
3610 struct ceph_string *pool_ns = NULL;
3611 int mds = session->s_mds;
3614 struct ceph_vino vino;
3616 u64 inline_version = 0;
3617 void *inline_data = NULL;
3620 size_t snaptrace_len;
3623 dout("handle_caps from mds%d\n", mds);
3626 end = msg->front.iov_base + msg->front.iov_len;
3627 tid = le64_to_cpu(msg->hdr.tid);
3628 if (msg->front.iov_len < sizeof(*h))
3630 h = msg->front.iov_base;
3631 op = le32_to_cpu(h->op);
3632 vino.ino = le64_to_cpu(h->ino);
3633 vino.snap = CEPH_NOSNAP;
3634 seq = le32_to_cpu(h->seq);
3635 mseq = le32_to_cpu(h->migrate_seq);
3638 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3639 p = snaptrace + snaptrace_len;
3641 if (le16_to_cpu(msg->hdr.version) >= 2) {
3643 ceph_decode_32_safe(&p, end, flock_len, bad);
3644 if (p + flock_len > end)
3649 if (le16_to_cpu(msg->hdr.version) >= 3) {
3650 if (op == CEPH_CAP_OP_IMPORT) {
3651 if (p + sizeof(*peer) > end)
3655 } else if (op == CEPH_CAP_OP_EXPORT) {
3656 /* recorded in unused fields */
3657 peer = (void *)&h->size;
3661 if (le16_to_cpu(msg->hdr.version) >= 4) {
3662 ceph_decode_64_safe(&p, end, inline_version, bad);
3663 ceph_decode_32_safe(&p, end, inline_len, bad);
3664 if (p + inline_len > end)
3670 if (le16_to_cpu(msg->hdr.version) >= 5) {
3671 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3674 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3675 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3678 if (le16_to_cpu(msg->hdr.version) >= 8) {
3680 u32 caller_uid, caller_gid;
3684 ceph_decode_64_safe(&p, end, flush_tid, bad);
3686 ceph_decode_32_safe(&p, end, caller_uid, bad);
3687 ceph_decode_32_safe(&p, end, caller_gid, bad);
3689 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3690 if (pool_ns_len > 0) {
3691 ceph_decode_need(&p, end, pool_ns_len, bad);
3692 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3698 inode = ceph_find_inode(sb, vino);
3699 ci = ceph_inode(inode);
3700 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3703 mutex_lock(&session->s_mutex);
3705 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3709 dout(" i don't have ino %llx\n", vino.ino);
3711 if (op == CEPH_CAP_OP_IMPORT) {
3712 cap = ceph_get_cap(mdsc, NULL);
3713 cap->cap_ino = vino.ino;
3714 cap->queue_release = 1;
3715 cap->cap_id = le64_to_cpu(h->cap_id);
3718 cap->issue_seq = seq;
3719 spin_lock(&session->s_cap_lock);
3720 list_add_tail(&cap->session_caps,
3721 &session->s_cap_releases);
3722 session->s_num_cap_releases++;
3723 spin_unlock(&session->s_cap_lock);
3725 goto flush_cap_releases;
3728 /* these will work even if we don't have a cap yet */
3730 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3731 handle_cap_flushsnap_ack(inode, tid, h, session);
3734 case CEPH_CAP_OP_EXPORT:
3735 handle_cap_export(inode, h, peer, session);
3738 case CEPH_CAP_OP_IMPORT:
3740 if (snaptrace_len) {
3741 down_write(&mdsc->snap_rwsem);
3742 ceph_update_snap_trace(mdsc, snaptrace,
3743 snaptrace + snaptrace_len,
3745 downgrade_write(&mdsc->snap_rwsem);
3747 down_read(&mdsc->snap_rwsem);
3749 handle_cap_import(mdsc, inode, h, peer, session,
3751 handle_cap_grant(mdsc, inode, h, &pool_ns,
3752 inline_version, inline_data, inline_len,
3753 msg->middle, session, cap, issued);
3755 ceph_put_snap_realm(mdsc, realm);
3759 /* the rest require a cap */
3760 spin_lock(&ci->i_ceph_lock);
3761 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3763 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3764 inode, ceph_ino(inode), ceph_snap(inode), mds);
3765 spin_unlock(&ci->i_ceph_lock);
3766 goto flush_cap_releases;
3769 /* note that each of these drops i_ceph_lock for us */
3771 case CEPH_CAP_OP_REVOKE:
3772 case CEPH_CAP_OP_GRANT:
3773 __ceph_caps_issued(ci, &issued);
3774 issued |= __ceph_caps_dirty(ci);
3775 handle_cap_grant(mdsc, inode, h, &pool_ns,
3776 inline_version, inline_data, inline_len,
3777 msg->middle, session, cap, issued);
3780 case CEPH_CAP_OP_FLUSH_ACK:
3781 handle_cap_flush_ack(inode, tid, h, session, cap);
3784 case CEPH_CAP_OP_TRUNC:
3785 handle_cap_trunc(inode, h, session);
3789 spin_unlock(&ci->i_ceph_lock);
3790 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3791 ceph_cap_op_name(op));
3798 * send any cap release message to try to move things
3799 * along for the mds (who clearly thinks we still have this
3802 ceph_send_cap_releases(mdsc, session);
3805 mutex_unlock(&session->s_mutex);
3808 ceph_put_string(pool_ns);
3812 pr_err("ceph_handle_caps: corrupt message\n");
3818 * Delayed work handler to process end of delayed cap release LRU list.
3820 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3822 struct inode *inode;
3823 struct ceph_inode_info *ci;
3824 int flags = CHECK_CAPS_NODELAY;
3826 dout("check_delayed_caps\n");
3828 spin_lock(&mdsc->cap_delay_lock);
3829 if (list_empty(&mdsc->cap_delay_list))
3831 ci = list_first_entry(&mdsc->cap_delay_list,
3832 struct ceph_inode_info,
3834 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3835 time_before(jiffies, ci->i_hold_caps_max))
3837 list_del_init(&ci->i_cap_delay_list);
3839 inode = igrab(&ci->vfs_inode);
3840 spin_unlock(&mdsc->cap_delay_lock);
3843 dout("check_delayed_caps on %p\n", inode);
3844 ceph_check_caps(ci, flags, NULL);
3848 spin_unlock(&mdsc->cap_delay_lock);
3852 * Flush all dirty caps to the mds
3854 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3856 struct ceph_inode_info *ci;
3857 struct inode *inode;
3859 dout("flush_dirty_caps\n");
3860 spin_lock(&mdsc->cap_dirty_lock);
3861 while (!list_empty(&mdsc->cap_dirty)) {
3862 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3864 inode = &ci->vfs_inode;
3866 dout("flush_dirty_caps %p\n", inode);
3867 spin_unlock(&mdsc->cap_dirty_lock);
3868 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3870 spin_lock(&mdsc->cap_dirty_lock);
3872 spin_unlock(&mdsc->cap_dirty_lock);
3873 dout("flush_dirty_caps done\n");
3876 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3879 int bits = (fmode << 1) | 1;
3880 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3881 if (bits & (1 << i))
3882 ci->i_nr_by_mode[i]++;
3887 * Drop open file reference. If we were the last open file,
3888 * we may need to release capabilities to the MDS (or schedule
3889 * their delayed release).
3891 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3894 int bits = (fmode << 1) | 1;
3895 spin_lock(&ci->i_ceph_lock);
3896 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3897 if (bits & (1 << i)) {
3898 BUG_ON(ci->i_nr_by_mode[i] == 0);
3899 if (--ci->i_nr_by_mode[i] == 0)
3903 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3904 &ci->vfs_inode, fmode,
3905 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3906 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3907 spin_unlock(&ci->i_ceph_lock);
3909 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3910 ceph_check_caps(ci, 0, NULL);
3914 * Helpers for embedding cap and dentry lease releases into mds
3917 * @force is used by dentry_release (below) to force inclusion of a
3918 * record for the directory inode, even when there aren't any caps to
3921 int ceph_encode_inode_release(void **p, struct inode *inode,
3922 int mds, int drop, int unless, int force)
3924 struct ceph_inode_info *ci = ceph_inode(inode);
3925 struct ceph_cap *cap;
3926 struct ceph_mds_request_release *rel = *p;
3930 spin_lock(&ci->i_ceph_lock);
3931 used = __ceph_caps_used(ci);
3932 dirty = __ceph_caps_dirty(ci);
3934 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3935 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3936 ceph_cap_string(unless));
3938 /* only drop unused, clean caps */
3939 drop &= ~(used | dirty);
3941 cap = __get_cap_for_mds(ci, mds);
3942 if (cap && __cap_is_valid(cap)) {
3944 ((cap->issued & drop) &&
3945 (cap->issued & unless) == 0)) {
3946 if ((cap->issued & drop) &&
3947 (cap->issued & unless) == 0) {
3948 int wanted = __ceph_caps_wanted(ci);
3949 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3950 wanted |= cap->mds_wanted;
3951 dout("encode_inode_release %p cap %p "
3952 "%s -> %s, wanted %s -> %s\n", inode, cap,
3953 ceph_cap_string(cap->issued),
3954 ceph_cap_string(cap->issued & ~drop),
3955 ceph_cap_string(cap->mds_wanted),
3956 ceph_cap_string(wanted));
3958 cap->issued &= ~drop;
3959 cap->implemented &= ~drop;
3960 cap->mds_wanted = wanted;
3962 dout("encode_inode_release %p cap %p %s"
3963 " (force)\n", inode, cap,
3964 ceph_cap_string(cap->issued));
3967 rel->ino = cpu_to_le64(ceph_ino(inode));
3968 rel->cap_id = cpu_to_le64(cap->cap_id);
3969 rel->seq = cpu_to_le32(cap->seq);
3970 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3971 rel->mseq = cpu_to_le32(cap->mseq);
3972 rel->caps = cpu_to_le32(cap->implemented);
3973 rel->wanted = cpu_to_le32(cap->mds_wanted);
3979 dout("encode_inode_release %p cap %p %s\n",
3980 inode, cap, ceph_cap_string(cap->issued));
3983 spin_unlock(&ci->i_ceph_lock);
3987 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3989 int mds, int drop, int unless)
3991 struct dentry *parent = NULL;
3992 struct ceph_mds_request_release *rel = *p;
3993 struct ceph_dentry_info *di = ceph_dentry(dentry);
3998 * force an record for the directory caps if we have a dentry lease.
3999 * this is racy (can't take i_ceph_lock and d_lock together), but it
4000 * doesn't have to be perfect; the mds will revoke anything we don't
4003 spin_lock(&dentry->d_lock);
4004 if (di->lease_session && di->lease_session->s_mds == mds)
4007 parent = dget(dentry->d_parent);
4008 dir = d_inode(parent);
4010 spin_unlock(&dentry->d_lock);
4012 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4015 spin_lock(&dentry->d_lock);
4016 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4017 dout("encode_dentry_release %p mds%d seq %d\n",
4018 dentry, mds, (int)di->lease_seq);
4019 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4020 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4021 *p += dentry->d_name.len;
4022 rel->dname_seq = cpu_to_le32(di->lease_seq);
4023 __ceph_mdsc_drop_dentry_lease(dentry);
4025 spin_unlock(&dentry->d_lock);